Skip to content

Releases

Ponko v2.2.0-beta.1

Ponko v2.2.0-beta.1 Pre-release
Pre-release

Choose a tag to compare

@visitor257 visitor257 released this 19 Nov 12:44

Ponko is an offline-capable AI app for Android: chat (image, file and document input) +
drawing + tagging, each of which can be served over its own OpenAI-compatible API.
All inference runs on-device: apart from the optional in-app LoRA download, no feature
needs the network. No telemetry; your data stays on your phone.
See the README.

This is a pre-release. It installs alongside the stable app (package
com.litertchat.app.beta).

What changed since 2.2.0-beta.0

Sampling cache (EasyCache / UCache)

Two new rows in the parameter card of the drawing and image-to-image pages: a cache mode
dropdown and a threshold field. Both are passed per image, so switching them does not
reload the model.

  • EasyCache targets DiT models (SD3.5, Flux, Qwen-Image, H3-Image...)
  • UCache targets UNet models (SD1.5, SDXL)
  • Picking the wrong one for the architecture is not fatal: sd.cpp logs that the cache is
    not supported for that model type and then behaves as if it were off
  • Threshold left empty means the mode's own default (EasyCache 0.2, UCache 1.0)

Reading the result: max_warmup_steps defaults to 8, so a run with fewer than ~20 steps
cannot skip anything by design. The engine states what it actually did in the log
(EasyCache skipped N/M steps (x.xx speedup)); a skipped step also reports progress, so
the bar keeps moving normally.

CLIP skip

A third new row. Empty means "use the architecture default": 1 for SD1, 2 for SD2 / SDXL /
SD3 / Flux. On SD1.5, an empty field and an explicit 1 must produce a pixel-identical
image with the same seed. Values are clamped to 1..12. It only affects encoders that have
layers to skip - the T5 and LLM encoders (SD3, Flux, Qwen, H3) ignore it.

Remaining time

The progress line used to count up ("· 1234 s"), which says nothing about how much longer
you are waiting. It now ends with an estimate extrapolated from the average time per step.

  • The unit is a setting: About → Settings → Remaining time, seconds or minutes,
    seconds by default
  • It appears from the first completed step; the text encoding before it is not
    proportional to steps and is excluded from the timing
  • Switching the unit takes effect immediately on a running job, without interrupting it

Load failures now carry the reason

"Load failed: could not create inference context. The file may be unsupported..." was all
the app said when sd.cpp refused a file - which cannot be acted on. sd.cpp does log the
real reason (a tensor it could not find, a shape that did not match, a header it could not
read), but that only reached the log file. The most informative line is now appended to
the message itself, so a screenshot carries the cause. When no such line exists the
message is unchanged.

Fixed: loading progress was reported as sampling steps

The bar could sit at "step 528/528", full and frozen, with no estimate. 528 was never a
step count - it is how many tensors the text encoder loads. The engine latches "sampling
started" the first time it reports a step count equal to the one requested, and a small
component inside the H3-Image text encoder happens to report exactly 6 tensors, which was
the requested step count. Every tensor batch after that was drawn as a sampling step, and
because total == current the estimate was suppressed.

The guard now also requires the reported total to equal the requested steps. A tensor
batch never matches; real sampling always does.

Fixed: the elapsed time and the estimate kept disappearing

The progress line is written from two places - a one-second ticker, and the preview-frame
callback that fires once per sampling step. The ticker wrote the full line; the preview
callback wrote only "Denoising… step N/M" and knew nothing about the start time.

Whoever wrote last won. On H3-Image, which emits a preview frame on every step, the
elapsed time and the estimate vanished a moment after each tick and came back a second
later. Both paths now build the line through the same function.

Verified on a device

512×512, 6 steps, CFG 1.0, Euler + Discrete, minimax_h3_image_pruned-Q2_K on a 15.6 GB
phone:

  • the line starts at step 0/6 with an empty bar - not 528/528
  • after the first step: step 1/6 · 273 s · ~1201 s left
  • switching the unit to minutes turns the same line into ~23 min while the job keeps
    running
  • at step 2/6 it still reads 574 s · ~1083 s left, and 25 seconds later
    600 s · ~1135 s left - the preview frames no longer wipe it

中文说明

Ponko 是一个可离线运行的 Android AI 应用:对话(支持图片、文件、文档输入)+ 绘图 + 打标,
三者都能各自开一个 OpenAI 兼容 API。所有推理都在设备本地完成——除了可选的 App 内 LoRA
下载,任何功能都不需要联网。无遥测,数据留在你手机上。

这是预发布版。 它与正式版并存(包名 com.litertchat.app.beta)。

相对 2.2.0-beta.0 改了什么

采样缓存(EasyCache / UCache)

绘图页和重绘页的参数卡里多了两行:缓存模式下拉框 + 阈值输入框。两个都是每张图现传的,
切换不会重载模型。

  • EasyCache 面向 DiT 模型(SD3.5、Flux、Qwen-Image、H3-Image 等)
  • UCache 面向 UNet 模型(SD1.5、SDXL)
  • 选错不致命:sd.cpp 会记一句「该模型类型不支持此缓存」,然后当作没开
  • 阈值留空 = 按模式取默认值(EasyCache 0.2,UCache 1.0)

怎么看有没有生效:max_warmup_steps 默认 8,所以 20 步以下的运行本来就不会跳步。
引擎会把实际做了什么写进日志(EasyCache skipped N/M steps (x.xx speedup));被跳过的步
同样会上报进度,进度条照常走。

CLIP skip

第三行。留空表示「用架构默认值」:SD1 是 1,SD2 / SDXL / SD3 / Flux 是 2。在 SD1.5 上,
留空和显式填 1 必须出一张同 seed 完全一致的图。取值夹在 1~12。它只对有层可跳的编码器
起作用——T5 和 LLM 类编码器(SD3、Flux、Qwen、H3)会忽略它。

剩余时间

进度行以前是往上涨的(「已 1234 秒」),读不出还要等多久。现在末尾会给出一个按「平均每步
耗时」外推的估算。

  • 单位是设置项:关于 → 设置 → 剩余时间单位,秒或分钟,默认秒
  • 从第一步完成起出现;首步之前的文本编码和步数不成比例,不计入计时
  • 出图途中切单位立刻生效,不会打断正在跑的任务

加载失败会带上原因

sd.cpp 拒绝一个文件时,以前只显示「加载失败:无法创建推理上下文。文件可能不受支持…」,
这种话没法照着做。sd.cpp 其实会打出真正的原因(哪个张量没找到、哪个形状不对、哪个头读不
出来),但那几行只进日志文件。现在最有信息量的那一行会拼进提示本身,截个图就够。没有这类
日志时,提示与以前完全一致。

修复:加载进度被当成采样步显示

进度条可能停在第 528/528 步、满格不动,而且一直不显示剩余时间。528 从来不是步数——那是
文本编码器加载的张量个数。引擎在「上报的步数 == 请求步数」时会锁存「已经开始采样」,而
H3-Image 文本编码器里有个小组件恰好上报 6 个张量,正好等于请求的步数。锁存之后,后面每一批
张量都被画成一步采样,又因为总数等于当前值,估算就整个被压掉了。

现在判据多了一条:上报的总数还必须等于请求的步数。张量批次永远对不上,真正的采样永远对
得上。

修复:耗时和剩余时间会突然消失

进度行是从两个地方写的——每秒一次的 ticker,和每步一次的预览帧回调。ticker 写完整的一行;
预览帧回调只写「正在去噪采样… 第 X/Y 步」,而且它拿不到开始时刻。

谁后写谁赢。H3-Image 每一步都出预览帧,于是耗时和剩余时间在每次 tick 后一闪就没了,一秒后又
回来。现在两条路径都走同一个函数生成那行字。

已在设备上验证

512×512、6 步、CFG 1.0、Euler + Discrete、minimax_h3_image_pruned-Q2_K,15.6 GB 手机:

  • 开跑是 第 0/6 步,进度条从空开始——不再是 528/528
  • 第一步完成后:第 1/6 步 · 已 273 秒 · 剩余约 1201 秒
  • 把单位切成分钟,同一行变成 剩余约 23 分钟,任务照常跑
  • 第 2 步时仍是完整的 已 574 秒 · 剩余约 1083 秒,25 秒后变成 已 600 秒 · 剩余约 1135 秒
    ——预览帧不再把它擦掉

Ponko v2.2.0-beta.0

Ponko v2.2.0-beta.0 Pre-release
Pre-release

Choose a tag to compare

@visitor257 visitor257 released this 16 Nov 08:36

Ponko is an offline-capable AI app for Android: chat (image, file and document input) +
drawing + tagging, each of which can be served over its own OpenAI-compatible API.
All inference runs on-device: apart from the optional in-app LoRA download, no feature
needs the network. No telemetry; your data stays on your phone.
See the README.

This is a pre-release. It installs alongside the stable app (package
com.litertchat.app.beta).

What changed since 2.1.3

stable-diffusion.cpp is upgraded to upstream master (2026-11-13) - 79 upstream commits,
ggml included.

That brings MiniMax-H3-Image and P-H3-Image (#2324, #2272). H3 video was already
there; what was missing was the Sol-H3 VAE decode path (#2317). CHROMA also gets its own
version id (#2237) instead of borrowing another model's.

Fixes in there that reach features which already worked: hires now honours the step count
you ask for (#2227), diffusers-named SDXL is recognised without the ComfyUI prefix (#2247),
LoRA scalars stay F32 and are merged as F32 when needed (#2263, #2302), and Krea2 Kohya LoRA
names no longer vanish (#2301).

  • New family MiniMax-H3-Image / P-H3-Image: diffusion model + Qwen3-VL encoder +
    Sol-H3 VAE. Reference-image editing works from the same checkpoint
  • New Text connector slot, needed when pairing the 4B encoder. Appended at the end of
    the slot table rather than next to llm, for the same reason the tokenizer slot was

Verified on a device

512×512, 6 steps, CFG 1.0, Euler + Discrete, shift left empty, on a 15.6 GB phone. 651
tensors loaded, the image comes out clean, no engine warnings.

  • minimax_h3_image_pruned-Q2_K.gguf - 7.2 GB
  • p-h3-image_text_encoder_int8_convrot.safetensors - 4.8 GB
  • minimax_h3_image_vae_fp16.safetensors - 1.1 GB

Please read before reporting

Expect about 25 minutes for one 512×512 image. 13.4 GB of weights on a 15.6 GB device
means the diffusion steps are read from storage rather than computed - the log shows
available memory falling to 1.2 GB. Slow, not broken.

Two things that waste time if you get them wrong:

  • Use the P-H3-Image text encoder, not the 32B one. 4.8 GB instead of 17 GB, and it
    handles both text-to-image and reference-image editing
  • Use the MiniMax-H3-Image VAE. The MiniMax-H3 video VAE and the P-H3-Image VAE are
    both incompatible and fail at decode

More steps do not help either; the model is trained on a six-step schedule.

中文说明

Ponko 是一个可离线运行的 Android AI 应用:对话(支持图片、文件、文档输入)+ 绘图 + 打标,
三者都能各自开一个 OpenAI 兼容 API。所有推理都在设备本地完成——除了可选的 App 内 LoRA
下载,任何功能都不需要联网。无遥测,数据留在你手机上。

这是预发布版。 它与正式版并存(包名 com.litertchat.app.beta)。

相对 2.1.3 改了什么

stable-diffusion.cpp 升级到上游 2026-11-13 的 master——79 个上游提交,ggml 一并跟上。

由此带来 MiniMax-H3-Image 和 P-H3-Image(#2324、#2272)。H3 视频本来就有,
缺的是 Sol-H3 VAE 的解码路径(#2317)。CHROMA 也终于有了自己的版本 id(#2237),
不再借用别人的。

其中对已有功能有实质影响的修复:hires 现在遵守你申请的步数(#2227)、
diffusers 命名的 SDXL 不带 ComfyUI 前缀也能识别(#2247)、
LoRA 标量保持 F32 并在需要时按 F32 合并(#2263、#2302)、
Krea2 的 Kohya LoRA 名字不再丢失(#2301)。

  • 新家族 MiniMax-H3-Image / P-H3-Image:扩散模型 + Qwen3-VL 编码器 + Sol-H3 VAE。
    同一个 checkpoint 也支持参考图编辑
  • 新增 Text connector 槽位,配 4B 编码器时需要。追加在槽位表末尾而不是放在 llm
    旁边,理由和当初的 tokenizer 槽位一样

已在设备上验证

512×512、6 步、CFG 1.0、Euler + Discrete、shift 留空,在 15.6 GB 手机上:
651 个张量加载成功,出图正常,引擎零警告。

  • minimax_h3_image_pruned-Q2_K.gguf — 7.2 GB
  • p-h3-image_text_encoder_int8_convrot.safetensors — 4.8 GB
  • minimax_h3_image_vae_fp16.safetensors — 1.1 GB

报告问题前请看

一张 512×512 预计要 25 分钟。 13.4 GB 权重放在 15.6 GB 内存的设备上,扩散的每一步
都是从存储里读数而不是在算——日志里可用内存掉到 1.2 GB。是慢,不是坏。

两个搞错就白等的地方:

  • 文本编码器用 P-H3-Image 那个,不要用 32B 的。 4.8 GB 对 17 GB,且文生图和参考图
    编辑都能用
  • VAE 必须用 MiniMax-H3-Image 的。 MiniMax-H3 的视频 VAE 和 P-H3-Image 的 VAE
    都不兼容,会在解码阶段失败

加步数也没用——这个模型是按六步调度的。

Ponko v2.1.3

Choose a tag to compare

@visitor257 visitor257 released this 12 Nov 09:40

中文说明  ·  跳到末尾  ·  Jump to end

Ponko is an offline-capable AI app for Android: chat (image, file and document input) +
drawing + tagging, each of which can be served over its own OpenAI-compatible API.
All inference runs on-device: apart from the optional in-app LoRA download, no feature
needs the network. No telemetry; your data stays on your phone.
See the README.

Installs over 2.1.2. Models, settings and LoRA selections carry across.

What's new

Flow shift is adjustable

Requested in issue #4. sd.cpp gives every flow-matching model family a default timestep shift -
FLUX.3 Image Klein 6.93, MOSS-Image 5.0, Fibo 2.0 3.0, Krea 2 1.15, Ideogram 4 and FLUX.2 Klein
1.0 - and we never passed the parameter, so those defaults were the only values reachable. When
a default does not suit the resolution being rendered there was nothing to turn.

Both parameter cards now have a Shift field. Empty means use the model's own default;
filling it overrides. The value is recorded in the draw log either way, so a run can be told
apart from one where it was left alone.

One thing to know: this only affects flow-matching models. SD1.5, SDXL, PixArt and other
epsilon-prediction models route through a different denoiser that never reads the value, so the
field is a no-op for them. That is why it is not hidden per model - deciding which families are
flow-based would be a second list to keep in sync with sd.cpp.

Fixed: diagnostic reports were losing the render log

Reported logs were missing render start, the parameter lines, every step and the result, while
keeping hundreds of unused lora tensor warnings. The cause was the 6000-character tail window
plus sd.cpp emitting one warning per mismatched tensor - several hundred of them, within five
milliseconds, all at the end of the file. They filled the window.

Runs of three or more lines sharing a signature are now collapsed before truncating. The
signature ignores the timestamp, anything inside |...|, and the trailing --- file:line, so
warnings collapse while step 3/20 and step 4/20 stay distinct. No keyword list to maintain.

Fixed: a loaded model could be leaked

sdHandle is per-Activity but the native context lives as long as the process. When Android
destroys and recreates the activity - language change, font scale, recovery after memory
pressure - the new page starts with no handle while the context is still alive, and nothing owns
it any more. Tapping load then built a second copy: two model sets of memory for one model.

On rebuild the page now takes the context back if the selected model set still matches, and frees
it if not. Reclaiming means no reload after, say, a language change. When the identity cannot be
determined it is freed rather than adopted - a possibly-wrong handle would crash mid-render,
which is worse than a reload.

Known limits

Unchanged from 2.1.0. The two configurations still untested are PixArt Alpha (the 512 variant)
and MOSS-Image at its recommended CFG.

中文说明

Ponko 是一个可离线运行的 Android AI 应用:对话(支持图片、文件、文档输入)+ 绘图 + 打标,
三者都能各自开一个 OpenAI 兼容 API。所有推理都在设备本地完成——除了可选的 App 内 LoRA
下载,任何功能都不需要联网。无遥测,数据留在你手机上。

可直接覆盖 2.1.2 安装。模型、设置和 LoRA 选择都会保留。

新增:可调 flow shift

来自 #4 的请求。sd.cpp 给每个 flow-matching
模型族都配了默认的时间步偏移——FLUX.3 Image Klein 6.93、MOSS-Image 5.0、Fibo 2.0 3.0、
Krea 2 1.15、Ideogram 4 与 FLUX.2 Klein 1.0——而我们从来没传过这个参数,所以那些默认值就是
唯一能用的值。默认值不适合当前分辨率时,没有任何可调的地方。

两个参数卡现在都有 Shift 输入框。留空 = 用模型自带的默认值,填了才覆盖。两种情况都会
写进绘图日志,所以事后能分清某一次到底有没有动过它。

有一点要说清楚:它只影响 flow 系模型。SD1.5、SDXL、PixArt 以及其它 EPS 预测的模型走的是另一个
denoiser,根本不读这个值,对它们填了也没用。之所以不按模型隐藏这个框,是因为「哪些族是 flow」
会变成一份需要跟 sd.cpp 保持同步的清单,两边不一致的代价比留一个空操作更大。

修复:诊断报告会丢掉出图日志

用户反馈的报告里,render start、参数行、每一步进度和结果全都不见了,只剩几百条
unused lora tensor 警告。原因是每份日志只留 6000 字符的尾部,而 sd.cpp 对每个对不上的张量
各打一行警告——几百行、5 毫秒内打完、全在文件末尾,正好把这 6000 字符占满。

现在截尾前会先折叠「连续三行以上签名相同」的行。签名忽略时间戳、|...| 里的内容和结尾的
--- file:line,所以警告会折叠而 step 3/20、step 4/20 仍然各自保留。不用维护关键字表。

修复:已加载的模型可能泄漏

sdHandle 是 Activity 的字段,而 native context 与进程同寿命。Android 销毁重建 Activity 时
(换语言、改字体、内存紧张后恢复),新建的页面拿不到句柄,而 context 还活着——此时没有任何
对象拥有它。再点「加载」就会建第二份:一个模型占两份内存。

现在重建时若选中的模型集仍匹配,就把 context 接回来(所以换语言之后不用重新加载);不匹配则
释放。判断不出身份时一律释放——拿一个可能不对的句柄出图会崩,那比重加载更糟。

已知限制

与 2.1.0 相同。仍未验证的两项是 PixArt Alpha(512 变体)和 MOSS-Image 在其推荐 CFG 下的表现。

Ponko v2.1.2

Choose a tag to compare

@visitor257 visitor257 released this 08 Nov 08:27

中文说明  ·  跳到末尾  ·  Jump to end

Ponko is an offline-capable AI app for Android: chat (image, file and document input) +
drawing + tagging, each of which can be served over its own OpenAI-compatible API.
All inference runs on-device: apart from the optional in-app LoRA download, no feature
needs the network. No telemetry; your data stays on your phone.
See the README.

Installs over 2.1.1. Models, settings and LoRA selections carry across.

What's new

More than one LoRA at a time, each with its own weight

Requested in #4. The app applied exactly one
LoRA: you could keep any number of .safetensors in the list and switch between them, but
only the selected one was loaded, and the strength was a single value shared by whatever was
loaded.

The LoRA list is now multi-select, and each ticked entry shows its own weight. The txt2img and
img2img parameter cards grow one weight field per selected LoRA, labelled with the file
name - so three LoRAs means three rows, not one global slider that can only mean one thing.
Both pages edit the same values and re-sync when you switch between them.

Nothing in the engine had to change. sd.cpp already takes a list of LoRAs with a multiplier
each; the bridge from Kotlin to native was passing one. The work was plumbing, but with two
places worth being careful: paths and weights are collected in pairs so a skipped entry cannot
shift a weight onto its neighbour, and the path strings are materialised before any pointer
into them is taken.

Three details that are decisions rather than mechanics:

A weight of 0 drops that LoRA before it reaches the engine. sd.cpp would otherwise load the
entire file and multiply the result by zero.

Turning the LoRA switch on with nothing selected selects the first file, which is what the
old single-slot code did implicitly. "On but doing nothing" would have read as a regression.

An existing user's global strength is migrated onto the LoRA they had selected, so a tuned
0.8 does not silently become 1.0.

One expectation to set: two LoRAs at 1.0 usually fight rather than stack - styles cancel,
fine detail goes muddy. 0.5-0.8 each is the usual starting point. Every extra file also costs
memory and load time.

Known limits

Unchanged from 2.1.0 — see that release's notes. The two configurations still untested are
PixArt Alpha (the 512 variant) and MOSS-Image at its recommended CFG.

中文说明

Ponko 是一个可离线运行的 Android AI 应用:对话(支持图片、文件、文档输入)+ 绘图 + 打标,
三者都能各自开一个 OpenAI 兼容 API。所有推理都在设备本地完成——除了可选的 App 内 LoRA
下载,任何功能都不需要联网。无遥测,数据留在你手机上。

可直接覆盖 2.1.1 安装。模型、设置和 LoRA 选择都会保留。

新增:可同时挂多份 LoRA,各自独立权重

来自 #4 的请求。原来只挂一份:列表里可以放任意多个
.safetensors 并切换,但同时只有一个会加载,强度是全局一个值。

现在列表是多选,选中的条目各自显示自己的权重。文生图与图生图的参数卡里,每选中一份就多一行
对应的权重输入框
,标签是文件名——选三份就是三行,而不是一个只能表达一个值的全局滑杆。
两页编辑同一份数据,切页时自动同步。

引擎完全没改。sd.cpp 本来就接受「一组 LoRA、各自带 multiplier」,是我们的 Kotlin 到 native
那层只传了一个。所以这是接线工作,但有两处要小心:路径与权重必须成对收集(否则跳过的条目会让
权重错位到别人身上),以及字符串要先全部落地再取指针(vector 扩容会让指针失效)。

三处属于判断而非机械的部分:

权重填 0 会在下发引擎前就丢掉这一份。否则 sd.cpp 会把整个文件加载完再乘 0。

打开 LoRA 开关但一份都没选时,自动选第一个——老的单槽代码隐式就是这么做的,
改成"开着却没反应"会被当成功能坏了。

老用户那个全局强度会迁移到他当时选中的那一份上,调好的 0.8 不会悄悄变回 1.0。

一点预期要提前说:两份都设 1.0 通常是互相打架而不是叠加——画风互相抵消、细节变糊。
一般各自 0.5~0.8 起步。每多挂一份,内存和加载时间都会增加。

已知限制

与 2.1.0 相同,见那一版的说明。仍未验证的两项是 PixArt Alpha(512 变体)和 MOSS-Image
在其推荐 CFG 下的表现。

Ponko v2.1.1

Choose a tag to compare

@visitor257 visitor257 released this 03 Nov 17:06

Ponko is an offline-capable AI app for Android: chat (image, file and document input) +
drawing + tagging, each of which can be served over its own OpenAI-compatible API.
All inference runs on-device: apart from the optional in-app LoRA download, no feature
needs the network. No telemetry; your data stays on your phone.
See the README.

Installs over 2.1.0. Models and settings carry across.

What's new

A search button on the Ponko-EX page

The plugin list used to carry a full-width search box under the toolbar, always visible. With
one or two plugins installed that is a whole row of height spent on something you never
touch. It is now hidden by default and opened by a magnifier in the top-right corner, in the
same place the chat page keeps its gear and plus buttons.

Matching is case-insensitive and covers the display name, the package name and the manifest
display name, so typing voxcpm finds com.litertchat.ponko.voxcpm2 as well as the card
title.

Collapsing the search clears the query. An invisible box still filtering the list is how you
end up wondering where two of your plugins went.

The button only appears on the card list, not inside a plugin's own UI, since there are no
cards to search there.

Fixed: the plugin About dialog was unreadable in the light theme

Reported as dark text on a dark background. The cause was not the dialog itself:

onCreate             theme resolved, setTheme(light or dark)     once
in-app theme switch  palette re-resolved, UI rebuilt             no setTheme

So starting with the system in dark mode and then picking light inside the app left the
Activity on the dark theme while the palette had already gone light. A dialog created after
that took its background from the stale dark theme and its text colour from the new light
palette. The reverse arrangement — system light, app dark — was broken the same way and would
have shown light on light.

Fixed at both levels. Switching the theme in-app now re-applies the Activity theme, which
fixes every dialog created afterwards, not just this one. And the About dialog no longer
takes colours from the theme at all: background, title and body come from the palette, and
the title is an ordinary TextView rather than setTitle(), whose text colour the framework
theme controls.

Known limits

Unchanged from 2.1.0 — see that release's notes. The two configurations still untested are
PixArt Alpha (the 512 variant) and MOSS-Image at its recommended CFG.

中文说明

Ponko 是一个可离线运行的 Android AI 应用:对话(支持图片、文件、文档输入)+ 绘图 + 打标,
三者都能各自开一个 OpenAI 兼容 API。所有推理都在设备本地完成——除了可选的 App 内 LoRA
下载,任何功能都不需要联网。无遥测,数据留在你手机上。

可直接覆盖 2.1.0 安装。模型与设置都会保留。

新增:Ponko-EX 页的搜索按钮

插件列表原来在工具栏下面常驻一整条全宽搜索框。只装一两个插件时,那是白白占掉一行高度、
你永远不会去点的东西。现在默认收起,由右上角的放大镜打开——和对话页 ⚙/+ 在同一个位置。

匹配不分大小写,覆盖显示名、包名和 manifest 里的显示名,所以搜 voxcpm 既能看到卡片标题,
也能命中 com.litertchat.ponko.voxcpm2。

收起时会清空查询。一个看不见却还在过滤列表的输入框,正是「我的插件怎么少了两个」的来源。

按钮只在卡片列表页出现,插件自己的界面里没有卡片可搜,那里不显示。

修复:插件「关于」弹窗在亮主题下看不清

报告的现象是暗底暗字。问题不在弹窗本身:

onCreate           解析主题、setTheme(亮/暗)        只跑一次
App 内切主题        重新解析配色、重建界面            没有 setTheme

所以「启动时系统是深色、之后在 App 里选亮色」会让 Activity 停留在深色主题,而配色已经变亮。
此时创建的弹窗,背景取自陈旧的深色主题、文字颜色取自新的亮色配色。反过来(系统亮、App 暗)
同样坏,会表现为亮底亮字。

两层都修了。App 内切换主题现在会重新应用 Activity 主题,这修好的是之后创建的所有弹窗,
不止这一个。另外「关于」弹窗彻底不再从主题取色:背景、标题、正文全部来自配色表,
标题改成普通 TextView 而不是 setTitle()——后者的文字颜色由框架主题控制。

已知限制

与 2.1.0 相同,见那一版的说明。仍未验证的两项是 PixArt Alpha(512 变体)和 MOSS-Image
在其推荐 CFG 下的表现。

Ponko v2.1.0

Choose a tag to compare

@visitor257 visitor257 released this 01 Nov 17:26

Ponko is an offline-capable AI app for Android: chat (image, file and document input) +
drawing + tagging, each of which can be served over its own OpenAI-compatible API.
All inference runs on-device: apart from the optional in-app LoRA download, no feature
needs the network. No telemetry; your data stays on your phone.
See the README.

Installs over 2.0.4. Models and settings carry across — you do not need to re-import
anything.

What's new

The bundled sd.cpp moved to upstream master

The engine was a 2026-09-10 snapshot; it is now upstream master as of 2026-10-26, with ggml
synced to match. That is where most of this release comes from.

Reference images

The drawing page has a reference-image card: up to two images, and the model reads them
alongside the prompt rather than only the prompt.

Worth being precise about what this is, because it is easy to confuse with img2img and the
two do different things. An img2img starting image becomes the canvas — it is encoded,
noised to the strength you chose, and denoised from there, so the composition is fixed.
Reference images are separate tokens appended to the canvas tokens and held at timestep 0,
clean throughout, while the canvas itself starts from noise. Composition stays free; the
model copies style.

The card only appears when the loaded model accepts reference images. That is asked of the
engine rather than guessed from the family name, because FLUX only takes them in its Kontext
variants, and handing reference latents to a model that does not expect them changes the
graph it was built around. Switching to a model without support clears your picks.

This is what Krea2's krea2_turbo_style_reference LoRA needed, and it also opens FLUX
Kontext, Qwen-Image-Edit, Fibo 2, Longcat and MageFlow.

PixArt

Reported as "never generated anything or fails to load" in every import mode. The engine
already supported PixArt; the failure was naming, in two places at once.

A diffusers-exported transformer carries no prefix at all — the 603 tensors in
PixArt-Sigma-XL-2-256x256 read pos_embed.proj.weight,
transformer_blocks.0.attn1.to_q.weight — while every detection branch asked for
model.diffusion_model.*, so the file was classified as something else entirely. And the
name converter rewrote the diffusers names correctly but relied on a later generic map to add
the prefix, which only fires for transformer. / unet. / diffusion_model.; bare names
never matched.

Both are fixed, PixArt has its own family now, and the import dialog asks for the three files
it actually needs: diffusion model, T5-XXL, VAE. It has no CLIP — caption_channels 4096 is
T5-XXL's hidden size.

Measured on a 15.6 GB device, CPU backend, 4 threads, at 256×256 with 20 steps and CFG 4.5:
591 seconds, lowest free memory 3238 MB.

Use the SDXL VAE. Sigma's weights carry no csize_embedder, so it keeps the 0.13025
scale factor, and an SD1.5 VAE gives wrong colours. The two files are 319.1 MB and 319.2 MB,
so size will not tell you which you have.

MOSS-Image

Same naming cause, plus a second one of its own. Comfy publishes its diffusion model in
diffusion_models/ with bare names — 379 tensors reading video_embedder.weight,
layers.0.proj.ffn_down.weight — and the detection branch asked for a prefixed name that was
not there.

That fixed, it still would not load: MOSS has no fallback tokenizer. Unlike the other
LLM-based encoders, which fall back to vocab tables compiled into the binary, it throws
outright without an external tokenizer.json. There is now a Tokenizer JSON slot in the
model set, and the import accepts a .json for that slot only — the slot is the one place
that takes text instead of weights.

The Qwen3.5 encoder is the expensive part. The diffusion model carries only a projection
layer; text encoding is a full 9B forward pass. A 2-bit GGUF (3 GB) works through the llm
slot and is much kinder to memory than the 7 GB w4a8.

Measured: 256×256, 20 steps, CFG 1.0, Euler/Discrete, 2517 seconds, lowest free memory
5191 MB, no warnings. The model card recommends 1024×1024 at 50 steps with guidance 3.0;
above CFG 1 the guidance does two transformer forwards per step, so that run would have taken
about 84 minutes here, and 1024×1024 is out of reach on CPU-only hardware.

Diagnostics

Two gaps that made earlier reports unreadable, both closed:

  • A failed load explained nothing. sd.cpp knows exactly what went wrong and logs it, but
    only the render path ever drained that buffer, so a load failure surfaced as
    new_sd_ctx returned 0 and no evidence. During context creation every line is now
    captured, and both outcomes drain it into the DRAW log
  • Reference images were invisible in logs. The two lines proving they reached the model
    are LOG_INFO and mention no LoRA, so a filter that only passed LoRA lines dropped them.
    A run with references and a run without produced identical reports. The app now records
    the count itself — refImages=1 800x600, or refImages=none

Also from the 2.0.x line

These shipped during the 2.0.x series and are included here:

  • Per-conversation system prompt, set from a gear button on the chat page. Saving applies
    it to the ongoing conversation instead of making you start over, and the transcript is
    carried across. Clearing is done by saving an empty field
  • Fixed: a prompt edited mid-conversation had no effect. It was being appended to the
    transcript rather than set as the system instruction
  • LoRA diagnostics: the applied-tensor count and the "skip incompatible LoRA tensor"
    notices now reach the DRAW log

Known limits

  • PixArt Alpha (the 512 variant) is untested. Only Sigma was verified
  • MOSS at its recommended settings is untested. The run above is 256×256 at CFG 1.0 —
    negative prompts have never been exercised on this model
  • Krea2 is slow on CPU. 512×512 at 8 steps took 91.7 minutes on the device above, and
    reference images are not resized: a 1024×1024 reference beside a 256×256 canvas is sixteen
    times the canvas cost
  • MOSS needs roughly 12 GB of weights in its smallest complete form

中文说明

Ponko 是一个可离线运行的 Android AI 应用:对话(支持图片、文件、文档输入)+ 绘图 + 打标,
三者都能各自开一个 OpenAI 兼容 API。所有推理都在设备本地完成——除了可选的 App 内 LoRA
下载,任何功能都不需要联网。无遥测,数据留在你手机上。

可直接覆盖 2.0.4 安装。模型与设置都会保留,不需要重新导入。

新增内容

内置 sd.cpp 升到 upstream master

引擎原本是 2026-09-10 的快照,现在是 upstream master(2026-10-26),ggml 同步对齐。
这一版的大部分内容都来自这次升级。

参考图

绘图页多了一张参考图卡片:最多两张,模型会连同提示词一起参考它们,而不只是看提示词。

这里要说准确,因为它和图生图很容易混,而两者做的事不一样。图生图的起始图就是画布——
它被编码、按你选的 strength 加噪、再从那里去噪,构图因此是固定的。参考图则是作为独立 token
追加在画布 token 之后,并且始终保持在 timestep 0 的干净状态,画布本身从噪声开始。构图完全
自由,模型学的是风格。

卡片只在当前模型接受参考图时才出现。这个能力是问引擎的,不按家族名猜——因为 FLUX 只有
Kontext 变体接受参考图,而把参考图交给不认识它的模型会改变模型预期的图结构。换到不支持
参考图的模型会清空已选。

这就是 Krea2 的 krea2_turbo_style_reference LoRA 需要的东西,同时也解锁了 FLUX Kontext、
Qwen-Image-Edit、Fibo 2、Longcat 和 MageFlow。

PixArt

之前报告的现象是「放进去什么也生成不出来,或者加载失败」,三种导入方式都一样。引擎本来
就支持 PixArt,问题出在命名上,而且是两处同时出问题。

diffusers 直接导出的 transformer 没有任何前缀——PixArt-Sigma-XL-2-256x256 里 603 个
张量长这样:pos_embed.proj.weight、transformer_blocks.0.attn1.to_q.weight——而所有探测
分支都要求 model.diffusion_model.*,于是这个文件被当成了别的架构。同时,名字转换器虽然把
diffusers 的名字改对了,却依赖后面一张通用映射表补前缀,而那张表只认 transformer. /
unet. / diffusion_model.,裸名永远匹配不上。

两处都修了,PixArt 现在有独立家族,导入时会明确要求它真正需要的三个文件:扩散模型、
T5-XXL、VAE。它完全不用 CLIP——caption_channels 4096 就是 T5-XXL 的 hidden size。

在 15.6 GB 设备、CPU 后端、4 线程下实测:256×256、20 步、CFG 4.5,591 秒,
最低可用内存 3238 MB。

VAE 请用 SDXL 的。Sigma 的权重里没有 csize_embedder,所以保持 0.13025 的缩放因子,
用 SD1.5 的 VAE 会偏色。这两个文件一个是 319.1 MB、一个是 319.2 MB,光看大小分不出来。

MOSS-Image

同一类命名问题,加上它自己独有的第二个问题。Comfy 把扩散模型放在 diffusion_models/ 下,
张量是裸名——379 个,长这样:video_embedder.weight、layers.0.proj.ffn_down.weight——
而探测分支要的是带前缀的名字,自然找不到。

修完这个它还是加载不了:MOSS 没有内置词表回退。其他基于 LLM 的编码器会回退到编进二进制
的词表,它不会——没有外部 tokenizer.json 就直接抛异常。现在模型集里有了 Tokenizer JSON
槽位,导入时也只有这个槽位接受 .json——它是唯一一个吃文本而不是权重的槽位。

Qwen3.5 编码器才是最贵的部分。扩散模型里只有一个投影层,文本编码是一次完整的 9B 前向。
2-bit 的 GGUF(3 GB)能通过 llm 槽正常工作,对内存比 7 GB 的 w4a8 友好得多。

实测:256×256、20 步、CFG 1.0、Euler/Discrete,2517 秒,最低可用内存 5191 MB,
零警告。模型卡推荐的是 1024×1024、50 步、guidance 3.0;CFG 大于 1 时 guidance 每步要跑两遍
transformer,那样的运行在这台机器上约 84 分钟,而 1024×1024 在纯 CPU 上不现实。

诊断

两个让早期报告读不出信息的缺口,都补上了:

  • 加载失败什么也不说。 sd.cpp 清楚知道出了什么问题并且记了日志,但只有出图路径会取走
    那个缓冲,于是加载失败只表现为 new_sd_ctx returned 0,没有任何证据。现在上下文创建期间
    每一行都会被捕获,成败两种情况都会写进 DRAW 日志
  • 参考图在日志里看不见。 能证明参考图进到模型的那两行是 LOG_INFO 且不含 LoRA,
    而过滤器只放行 LoRA 行,于是被整条丢掉。「加了参考图」和「没加」两次运行产出完全相同的报告。
    现在 App 自己记录数量——refImages=1 800x600,或者 refImages=none

来自 2.0.x 的内容

以下在 2.0.x 系列里已经发布,此处一并列出:

  • 按对话保存的系统提示词,对话页齿轮按钮打开底部抽屉。保存后对当前对话立即生效,不用
    重新开对话,历史照常保留。「清空」改成保存空输入框
  • 修复:对话中途改提示词不生效。提示词当时被塞进对话历史而不是系统指令
  • LoRA 诊断:应用到的张量数和「跳过不兼容的 LoRA 张量」提示,现在都会进 DRAW 日志

已知限制

  • PixArt Alpha(512 变体)未测试。 只验证了 Sigma
  • MOSS 在推荐参数下未测试。 上面的运行是 256×256、CFG 1.0——这个模型的负向提示词
    从未跑过
  • Krea2 在 CPU 上很慢。 上面的设备上 512×512、8 步花了 91.7 分钟;而且参考图不会被缩放,
    256×256 画布旁边放一张 1024×1024 参考图,代价是画布的十六倍
  • MOSS 最小完整组合约需 12 GB 权重

Ponko v2.1.0-beta.5

Ponko v2.1.0-beta.5 Pre-release
Pre-release

Choose a tag to compare

@visitor257 visitor257 released this 01 Nov 11:00

Ponko is an offline-capable AI app for Android: chat (image, file and document input) +
drawing + tagging, each of which can be served over its own OpenAI-compatible API.
All inference runs on-device: apart from the optional in-app LoRA download, no feature
needs the network. No telemetry; your data stays on your phone.
See the README.

This is a pre-release. It installs alongside the stable app (package
com.litertchat.app.beta).

What changed since 2.1.0-beta.4

MOSS-Image can be given its tokenizer, which it cannot load without.

beta.3 fixed the tensor-name prefix that stopped PixArt and MOSS from loading at all. For
MOSS that was necessary but not sufficient. Its conditioner is an LLM encoder, and unlike
the other LLM families - which fall back to the vocab tables compiled into the binary -
MOSS has no fallback:

if (version == VERSION_MOSS_IMAGE) {
    throw std::runtime_error("MOSS-Image requires the official Qwen3.5 tokenizer.json");
}

It needs tokenizer.json supplied from outside. The app had no way to supply one: the
field was never wired, no slot existed, and the slot table ended at TAESD.

  • New Tokenizer JSON slot. Appended at the end of the slot table rather than placed
    next to llm, because slot positions are persisted by index in existing model sets -
    inserting in the middle would shift everything after it and quietly mis-assign files
  • Import accepts .json for that slot only. The import path only took
    gguf/safetensors/ckpt/pt, so a .json was rejected before it could reach any slot.
    Picking a .json for a weights slot still fails, and with a message that says which
    mistake you made
  • MOSS gets its own slot template instead of sharing the LLM-encoder list, since its
    tokenizer is required rather than optional

Same slot also unblocks Lens and PiD, which hit the identical requirement in sd.cpp.

Still true from earlier betas

  • PixArt works (beta.3 + beta.4): 256×256, 20 steps, CFG 4.5, Euler/Discrete, 591 s on
    a 15.6 GB device, preview included. Use an SDXL VAE - Sigma keeps the 0.13025 scale
    factor and an SD1.5 VAE gives wrong colours
  • Reference images (beta.1/beta.2): up to two, only for models that accept them. Opens
    Krea2 style reference, FLUX Kontext, Qwen-Image-Edit, Fibo 2, Longcat and MageFlow
  • Load diagnostics (beta.3): the load path now drains sd.cpp's log buffer on both
    success and failure, so a failed load explains itself
  • The 2.0.4 features: per-conversation system prompt, its systemInstruction fix, and
    the drawing diagnostics that surface the engine's LoRA decisions

Please read before reporting

MOSS has still never been loaded. This release only makes it possible to try. Two things
to know before you do:

  • You need the Qwen3.5 tokenizer.json — the one from the Qwen3.5 release, not Qwen3-VL
  • The smallest complete set is about 12 GB (moss_image_sft_int8_convrot 5.4 GB +
    qwen3.5_9b_w4a8_gptq 7.0 GB + flux2-vae_small 0.25 GB). Whether that fits depends on
    your device; on a 15.6 GB phone with 6 GB free it may run out of memory during text
    encoding, before any sampling starts

中文说明

Ponko 是一个可离线运行的 Android AI 应用:对话(支持图片、文件、文档输入)+ 绘图 + 打标,
三者都能各自开一个 OpenAI 兼容 API。所有推理都在设备本地完成——除了可选的 App 内 LoRA
下载,任何功能都不需要联网。无遥测,数据留在你手机上。

这是预发布版。 它与正式版并存(包名 com.litertchat.app.beta)。

相对 2.1.0-beta.4 改了什么

MOSS-Image 现在可以给它 tokenizer 了——没有它,MOSS 加载不了。

beta.3 修的是让 PixArt 和 MOSS 完全加载不了的前缀问题。对 MOSS 来说那是必要条件,不是
充分条件。它的 conditioner 是 LLM 编码器,而和其他 LLM 家族不同——其他家族会回退到编进
二进制里的词表——MOSS 没有回退:

if (version == VERSION_MOSS_IMAGE) {
    throw std::runtime_error("MOSS-Image requires the official Qwen3.5 tokenizer.json");
}

它必须由外部提供 tokenizer.json。而 App 之前没有任何途径提供:字段没接、槽位不存在、
槽位表停在 TAESD。

  • 新增 Tokenizer JSON 槽位。 追加在槽位表末尾,而不是放在 llm 旁边——已有模型集按
    位置索引存槽位,插在中间会让后面全部错位,静默地把文件配给错误的槽
  • 只有这个槽位接受 .json。 导入路径原本只收 gguf/safetensors/ckpt/pt,
    .json 在到达槽位之前就被拒了。给权重槽位选 .json 仍然会失败,但会给出说明
    是哪种错误的信息
  • MOSS 单独一个槽位模板,不再和其他 LLM 编码器共用,因为它的 tokenizer 是必填而不是可选

同一个槽位也顺带解锁了 Lens 和 PiD——它们在 sd.cpp 里撞的是同一个要求。

更早的 beta 里仍然有效的内容

  • PixArt 可用(beta.3 + beta.4):256×256、20 步、CFG 4.5、Euler/Discrete,
    15.6 GB 设备上 591 秒,预览也正常。VAE 请用 SDXL 的——Sigma 保持 0.13025 缩放因子,
    用 SD1.5 的会偏色
  • 参考图(beta.1/beta.2):最多两张,只在当前模型接受参考图时显示。由此解锁 Krea2 的
    style reference、FLUX Kontext、Qwen-Image-Edit、Fibo 2、Longcat、MageFlow
  • 加载诊断(beta.3):加载路径现在成败两种情况都会取走 sd.cpp 的日志缓冲,
    失败的加载会自己解释原因
  • 2.0.4 的功能:按对话保存的系统提示词、它的 systemInstruction 修复、
    以及把引擎的 LoRA 判断输出到日志的绘图诊断

报告问题前请看

MOSS 仍然从来没有被加载过。 这一版只是让它「有可能试」。动手前要知道两件事:

  • 需要 Qwen3.5 的 tokenizer.json——是 Qwen3.5 那版,不是 Qwen3-VL 的
  • 最小的完整组合约 12 GB(moss_image_sft_int8_convrot 5.4 GB +
    qwen3.5_9b_w4a8_gptq 7.0 GB + flux2-vae_small 0.25 GB)。能不能放下取决于设备;
    在 15.6 GB、可用 6 GB 的手机上,可能在文本编码阶段就内存不足,采样还没开始

Ponko v2.1.0-beta.4

Ponko v2.1.0-beta.4 Pre-release
Pre-release

Choose a tag to compare

@visitor257 visitor257 released this 01 Nov 10:17

Ponko is an offline-capable AI app for Android: chat (image, file and document input) +
drawing + tagging, each of which can be served over its own OpenAI-compatible API.
All inference runs on-device: apart from the optional in-app LoRA download, no feature
needs the network. No telemetry; your data stays on your phone.
See the README.

This is a pre-release. It installs alongside the stable app (package
com.litertchat.app.beta).

What changed since 2.1.0-beta.3

PixArt works now, and its preview does too.

beta.3 fixed why PixArt would not load - a prefix applied twice, 603 of 603 tensors
unreachable. It loads and renders correctly now: 256×256, 20 steps, CFG 4.5,
Euler/Discrete, 591 seconds, lowest free memory 3238 MB, no tensor errors at all.

That run also exposed a smaller gap. The preview pane stayed blank for the whole render,
and the engine was saying so:

W: No latent to RGB projection known for this model

preview_image() chooses a latent-to-RGB projection by channel count. The 4-channel branch
knew SDXL and SD1/SD2 and sent everything else to the warning-and-return. PixArt is
4-channel, so it never got one.

The fix is to use one of the two existing tables - but it has to be the right one. PixArt
Alpha-512 ships an SD1.5 VAE (scale 0.18215) and Sigma ships an SDXL VAE (0.13025); the
latent spaces differ, so the coefficients are not interchangeable, and using the wrong one
shows up as wrong colours - which reads as a broken model rather than a wrong preview.

Rather than branch on version names, this reads the scale factor off first_stage_model -
the very value configure_pixart_vae() already decided from whether csize_embedder is
present. One source of truth, so the preview cannot disagree with how the image was
actually encoded.

Still true from earlier betas

  • Reference images (beta.1/beta.2): up to two, shown only for models that accept them.
    Opens Krea2 style reference, FLUX Kontext, Qwen-Image-Edit, Fibo 2, Longcat and MageFlow.
    Verified on a real device: the Krea2 LoRA matches 512/512 and the output style follows the
    reference
  • PixArt and MOSS name fix (beta.3): converters receive bare names and must return bare
    names. The wrapper appended a prefix the caller had already stripped
  • Load diagnostics (beta.3): the load path never drained sd.cpp's log buffer, so a
    failure produced one line and no evidence. During nativeCreate every INFO is captured,
    the cap rises to 200 lines, and the app drains it on both outcomes
  • The 2.0.4 features: per-conversation system prompt, its systemInstruction fix, and
    the drawing diagnostics that surface the engine's LoRA decisions

Please read before reporting

MOSS-Image still has never been loaded. It got the same prefix fix as PixArt, verified
the same way, but nobody has tried it - and it needs roughly 12 GB of weights, beyond most
phones.

中文说明

Ponko 是一个可离线运行的 Android AI 应用:对话(支持图片、文件、文档输入)+ 绘图 + 打标,
三者都能各自开一个 OpenAI 兼容 API。所有推理都在设备本地完成——除了可选的 App 内 LoRA
下载,任何功能都不需要联网。无遥测,数据留在你手机上。

这是预发布版。 它与正式版并存(包名 com.litertchat.app.beta)。

相对 2.1.0-beta.3 改了什么

PixArt 能用了,预览也正常了。

beta.3 修的是 PixArt 为什么加载不了——前缀被加了两次,603 个张量全部对不上。现在它加载
和出图都正确:256×256、20 步、CFG 4.5、Euler/Discrete,591 秒,最低可用内存
3238 MB,没有任何张量报错。

那次运行还暴露出一个小缺口:整个渲染过程预览框都是空的,而引擎其实一直在说原因:

W: No latent to RGB projection known for this model

preview_image() 按 latent 通道数挑投影表。4 通道那条分支只认 SDXL 和 SD1/SD2,
其余全部落到「警告并返回」。PixArt 就是 4 通道,所以一张都没拿到。

修法是从两张现成的表里选一张——但必须选对。PixArt Alpha-512 用的是 SD1.5 VAE
(scale 0.18215),Sigma 用的是 SDXL VAE(0.13025);两者 latent 空间不同,系数不能混用,
用错了表现为颜色不对——那看起来像是模型坏了,而不是预览错了。

所以这里不按版本名分支,而是直接读 first_stage_model 上的缩放因子——那正是
configure_pixart_vae() 按有没有 csize_embedder 定下来的值。单一事实来源,
预览不可能和实际编码方式对不上。

更早的 beta 里仍然有效的内容

  • 参考图(beta.1/beta.2):最多两张,只在当前模型接受参考图时显示。由此解锁 Krea2 的
    style reference、FLUX Kontext、Qwen-Image-Edit、Fibo 2、Longcat、MageFlow。
    已在真机验证:Krea2 的 LoRA 512/512 全部匹配,出图风格确实跟随参考图
  • PixArt 与 MOSS 的命名修复(beta.3):转换函数拿到的是裸名,也必须返回裸名。原包装补了
    一个调用方已经剥掉的前缀
  • 加载诊断(beta.3):加载路径从来不取 sd.cpp 的日志缓冲,所以失败只留下一行字、
    没有任何证据。现在 nativeCreate 期间全量捕获 INFO,上限提到 200 行,
    成败两种情况 App 都会取走
  • 2.0.4 的功能:按对话保存的系统提示词、它的 systemInstruction 修复、
    以及把引擎的 LoRA 判断输出到日志的绘图诊断

报告问题前请看

MOSS-Image 仍然从来没有被加载过。 它和 PixArt 拿到了同一处前缀修复,同样的验证方式,
但没人试过——而且它需要约 12 GB 权重,超出多数手机的范围。

Ponko v2.1.0-beta.3

Ponko v2.1.0-beta.3 Pre-release
Pre-release

Choose a tag to compare

@visitor257 visitor257 released this 01 Nov 04:35

Ponko is an offline-capable AI app for Android: chat (image, file and document input) +
drawing + tagging, each of which can be served over its own OpenAI-compatible API.
All inference runs on-device: apart from the optional in-app LoRA download, no feature
needs the network. No telemetry; your data stays on your phone.
See the README.

This is a pre-release. It installs alongside the stable app (package
com.litertchat.app.beta).

What changed since 2.1.0-beta.2

beta.2 claimed PixArt and MOSS-Image now load. Neither did. The cause was mine, in the code
I added for them, and it is fixed here.

The double prefix

Tensor names arrive at the converter already handled in both directions:

name = convert_diffusion_model_name(name.substr(prefix.size()), prefix, version);
name = prefix + name;   // the caller re-adds it

So converters work on bare names and must return bare names. I added a wrapper that
appended model.diffusion_model. whenever the result had no known top-level prefix - which
is always the case for a diffusion tensor, since that prefix is exactly what the caller
just stripped. Every one of the 603 PixArt tensors came out as
model.diffusion_model.model.diffusion_model.x_embedder.proj.weight. The model found none
of them and init() returned false, so the app could only report
new_sd_ctx returned 0.

MOSS carried the same wrapper with the same bug; it simply had not been loaded yet, so
PixArt reported it first. Both are fixed by reducing the wrapper to what it should have
been - stripping an optional transformer. layer and nothing else.

Verified by replaying the actual PixArt file through the actual call shape: 603 tensors in,
603 double-prefixed before, zero after, and every layer PixArt asks for resolves.

The silence that made it hard to see

sd.cpp knew what went wrong and logged it. The load path never drained that buffer - only
the render path called takeWarnings() - so a load failure produced exactly one line and
no evidence.

During nativeCreate the log filter now captures INFO as well as WARN/ERROR and the buffer
cap rises from 40 lines to 200. The app drains it on both outcomes: on failure everything
goes into the DRAW log, on success the INFO lines are dropped so render logs stay readable.
A side effect worth having - the load now records sd.cpp's Version: line, so whether
detection matched is visible rather than inferred.

Still true from beta.1 and beta.2

  • The 2.0.4 features: per-conversation system prompt, its systemInstruction fix, and the
    drawing diagnostics that surface the engine's LoRA decisions
  • Reference images: up to two, shown only for models that accept them. This opens Krea2
    style reference, FLUX Kontext, Qwen-Image-Edit, Fibo 2, Longcat and MageFlow
  • Krea2 style reference verified on a real device: the LoRA matches 512/512, and with a
    reference image the output style follows it

Please read before reporting

PixArt has still never loaded successfully. This release is the fix for why, verified
by replay against the real weight file, but not by generating an image. If it fails again,
the DRAW log now carries sd.cpp's own reason - please include it.

MOSS-Image has never been loaded at all. Same fix, same caveat, and it needs roughly
12 GB of weights which is beyond most phones.

中文说明

Ponko 是一个可离线运行的 Android AI 应用:对话(支持图片、文件、文档输入)+ 绘图 + 打标,
三者都能各自开一个 OpenAI 兼容 API。所有推理都在设备本地完成——除了可选的 App 内 LoRA
下载,任何功能都不需要联网。无遥测,数据留在你手机上。

这是预发布版。 它与正式版并存(包名 com.litertchat.app.beta)。

相对 2.1.0-beta.2 改了什么

beta.2 说 PixArt 和 MOSS-Image 现在能加载了。两个都不能。原因出在我为它们加的那段代码里,
这一版修掉了。

前缀被加了两次

张量名到转换函数手里时,前缀已经由调用方处理过两个方向:

name = convert_diffusion_model_name(name.substr(prefix.size()), prefix, version);
name = prefix + name;   // 调用方自己加回来

所以转换函数拿到的是裸名,也必须返回裸名。我加的那层包装看到结果没有已知顶层前缀,就自己补了
一个 model.diffusion_model.——而扩散模型张量永远符合这个条件,因为那个前缀正是调用方刚
剥掉的东西。PixArt 的 603 个张量全变成了
model.diffusion_model.model.diffusion_model.x_embedder.proj.weight,模型一个都找不到,
init() 返回 false,App 只能报一句 new_sd_ctx returned 0。

MOSS 带着一模一样的包装和一模一样的 bug,只是还没被加载过,所以先由 PixArt 报出来。
两个都修了——把包装缩回它本该做的事:只剥一层可选的 transformer.,别的什么都不做。

验证方式是把真实的 PixArt 文件按真实的调用形态重放一遍:603 个张量,修前 603 个双前缀,
修后 0 个,PixArt 需要的每一层都能对上。

让它难以发现的沉默

sd.cpp 知道出了什么问题,也打了日志。但加载路径从来不把那个缓冲取走——只有出图路径会调用
takeWarnings()——所以一次加载失败只留下一行字,没有任何证据。

现在 nativeCreate 期间过滤器除了 WARN/ERROR 还收 INFO,缓冲上限从 40 行提到 200 行。
App 在两种结果下都会取走:失败时全部写进 DRAW 日志,成功时丢掉 INFO 行,让出图日志保持干净。
顺带一个好处:加载时会记下 sd.cpp 的 Version: 一行,版本探测有没有命中从此是看到的,不是猜的。

beta.1 / beta.2 里仍然有效的内容

  • 2.0.4 的功能:按对话保存的系统提示词、它的 systemInstruction 修复、以及把引擎的
    LoRA 判断输出到日志的绘图诊断
  • 参考图:最多两张,只在当前模型接受参考图时显示。由此解锁 Krea2 的 style reference、
    FLUX Kontext、Qwen-Image-Edit、Fibo 2、Longcat、MageFlow
  • Krea2 style reference 已在真机验证:LoRA 512/512 全部匹配,加参考图后出图风格确实跟随

报告问题前请看

PixArt 至今没有成功加载过一次。 这一版是对「为什么不行」的修复,靠重放真实权重文件验证,
没有靠出图验证。如果还是失败,DRAW 日志里现在有 sd.cpp 自己给出的原因,请一并附上。

MOSS-Image 从来没有被加载过。 同一处修复,同样的保留意见,而且它需要约 12 GB 权重,
超出多数手机的范围。

Ponko v2.1.0-beta.2

Ponko v2.1.0-beta.2 Pre-release
Pre-release

Choose a tag to compare

@visitor257 visitor257 released this 31 Oct 06:55

Ponko is an offline-capable AI app for Android: chat (image, file and document input) +
drawing + tagging, each of which can be served over its own OpenAI-compatible API.
All inference runs on-device: apart from the optional in-app LoRA download, no feature
needs the network. No telemetry; your data stays on your phone.
See the README.

This is a beta build. It installs alongside the stable app (package
com.litertchat.app.beta).

What changed since 2.1.0-beta.1

Only diagnostics. Nothing about drawing, chat or the reference-image path itself is
different - if beta.1 worked for you, beta.2 works the same way.

A beta.1 report from a Krea2 run could not answer the question it was collected for:
whether the reference images had reached the model at all. The two lines inside sd.cpp that
say so - Using 'krea2_ostris_edit' preset for reference images and EDIT mode - are both
LOG_INFO, and the filter only passed INFO lines mentioning LoRA. They were dropped, so a
run with reference images and a run without them produced indistinguishable reports.

  • That filter now passes any INFO line containing lora, preset, reference image or
    edit mode. It stays a whitelist rather than passing everything, because a mismatched
    LoRA can emit hundreds of skip notices and the report is capped at 40 lines
  • The app writes the reference-image count into the DRAW log itself -
    refImages=1 800x600, or refImages=none - next to the existing size, steps, cfg and
    LoRA lines. sd.cpp's wording is upstream's to change; a line the app writes is ours

The reason this is worth its own build: one Krea2 render at 512x512 with 8 steps took
91.7 minutes on the device that reported it. A test that ends without a definite answer
costs an hour and a half to repeat.

Still true from beta.1

  • The 2.0.4 features: per-conversation system prompt, its systemInstruction fix, and the
    drawing diagnostics that surface the engine's LoRA decisions
  • PixArt Sigma / Alpha now loads from diffusers-exported files (use the SDXL VAE;
    Sigma keeps the 0.13025 scale factor and an SD1.5 VAE gives wrong colours)
  • MOSS-Image now loads from the Comfy layout (Qwen3.5 text encoder + Flux2 VAE)
  • Reference images: up to two, shown only when the loaded model takes them, which opens
    Krea2 style reference, FLUX Kontext, Qwen-Image-Edit, Fibo 2, Longcat and MageFlow

Please read before reporting

Still no real device test beyond Krea2 text-to-image. That one passed: the model loads,
(512 / 512) LoRA tensors have been applied, and the output quality is good. The
reference-image path itself has not been exercised on a real model yet, which is exactly
what beta.2's logging is for.

中文说明

Ponko 是一个可离线运行的 Android AI 应用:对话(支持图片、文件、文档输入)+ 绘图 + 打标,
三者都能各自开一个 OpenAI 兼容 API。所有推理都在设备本地完成——除了可选的 App 内 LoRA
下载,任何功能都不需要联网。无遥测,数据留在你手机上。

这是测试版。 它与正式版并存(包名 com.litertchat.app.beta)。

相对 2.1.0-beta.1 改了什么

只有诊断。绘图、对话、参考图通路本身没有任何变化——beta.1 能用的,beta.2 一样用。

beta.1 那份 Krea2 诊断报告回答不了它本来该回答的问题:参考图到底有没有送到模型。sd.cpp
里说这件事的两行——Using 'krea2_ostris_edit' preset for reference images 和
EDIT mode——都是 LOG_INFO,而过滤器只放行提到 LoRA 的 INFO 行。它们被丢掉了,于是
「加了参考图」和「没加参考图」两次运行产出的报告完全无法区分。

  • 过滤器现在放行任何含 lora、preset、reference image 或 edit mode 的 INFO 行。
    仍然用白名单而不是全放,因为一个不匹配的 LoRA 能吐出几百条跳过提示,而报告上限是 40 行
  • App 自己把参考图数量写进 DRAW 日志——refImages=1 800x600,或者 refImages=none——
    就写在已有的尺寸、步数、cfg、LoRA 那几行旁边。sd.cpp 的措辞是上游的,App 自己写的行才是我们的

为什么值得单独出一版:在那台报告设备上,Krea2 一张 512×512、8 步的图花了 91.7 分钟。
一个得不出确定结论的测试,重做一次就是一个半小时。

beta.1 里仍然有效的内容

  • 2.0.4 的功能:按对话保存的系统提示词、它的 systemInstruction 修复、以及把引擎的
    LoRA 判断输出到日志的绘图诊断
  • PixArt Sigma / Alpha 现在能加载 diffusers 导出的文件(VAE 请用 SDXL 的,Sigma 保持
    0.13025 缩放因子,用 SD1.5 的会偏色)
  • MOSS-Image 现在能加载 Comfy 的布局(Qwen3.5 文本编码器 + Flux2 VAE)
  • 参考图:最多两张,只在当前模型接受参考图时显示,由此解锁 Krea2 的 style reference、
    FLUX Kontext、Qwen-Image-Edit、Fibo 2、Longcat、MageFlow

报告问题前请看

除了 Krea2 的文生图,仍然没有经过真机测试。 那一次通过了:模型能加载、
(512 / 512) LoRA tensors have been applied、出图质量很好。参考图这条通路本身还没在
真实模型上跑过——beta.2 的日志改动就是为此。