Repository navigation
Ponko v2.1.0
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 0and 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
areLOG_INFOand 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, orrefImages=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