BluePixel T8 v0.1.11 — Attention dtype compatibility

Handle attention mask dtype conflicts in BluePixel text conditioning.

• Align floating masks with the query dtype before SDPA dispatch.
• Keep native CUDA BF16 masks on their existing path. Retry only the specific mask dtype error with an FP32 mask when a backend internally promotes Q/K/V.
• Preserve boolean-mask semantics, unrelated errors and host precision settings.

修复BluePixel文本注意力中BF16掩码与FP32 query的同类精度冲突。更新后请完全重启ComfyUI，模型无需重新下载。仍只交付8个画布工作流。

282 tests pass, including 15 new mask regressions on CPU/CUDA. Native BF16 small-tensor results match the previous path exactly. Real ComfyUI Attention/RMSNorm integration was checked with a tiny CPU model; full-model generation was not repeated. The supplied cloud traceback identifies BluePixel TextEncode and the plugin's reference_sdpa; update the node code on the cloud instance and restart its ComfyUI process. The user's cloud instance was not directly retested.

BluePixel T8 v0.1.10 — Canvas workflows

Only editable ComfyUI canvas workflows are shipped. The eight API-format files and their dedicated submission tool have been removed; the generator now writes canvas JSON only.

• Download BluePixel-Canvas-Workflows.zip, extract it, and open 02_edit_1k_preview.json with Ctrl+O or drop the local JSON onto ComfyUI.
• Complete widget defaults, workflow identity and initial view metadata; correct the SaveImage output definition against the current ComfyUI node schema.
• Eight workflows pass the installed official frontend schema; 128 nodes and 166 connections pass real node-definition checks. Six canvas regression tests pass. Visual drag/drop was not exercised because the UI tool could not bind a stable browser window.

仅保留8个画布工作流，已删除API工作流。建议先下载画布专用ZIP，解压后用Ctrl+O打开1K预览JSON，选择模型并上传参考图。

Existing model files remain compatible. No complete model inference or 4K generation was repeated for this update. Planning and research files remain excluded.

BluePixel T8 v0.1.9

Protect model source files from batch writes after a seventh 20-round review.

Batch logs could previously truncate a model file when the output directory and metadata filename targeted that file. The CLI could still report success. Model sources are now included in path and inode collision checks for metadata, PNGs, internal state and configuration writes. The CLI rejects conflicts before loading models; execution and graph entry points also protect sources retained by native clones.

第七次20轮复查完成，修复批任务日志或其它写入目标可能破坏模型源文件的问题，涵盖直接路径、硬链接及空计划配置写入。新增25项回归，261项测试全部通过。

Validation used temporary files, simulated sampling, real CLI/task storage and ComfyUI CPU node entry points. The suite includes the existing micro CUDA RoPE check. Full-model 4K generation was not repeated. Existing model files and valid task identities remain compatible. Restart ComfyUI after updating.

BluePixel T8 v0.1.8

Fix two issues confirmed during a sixth 20-round review:

• Preserve artifact identity through ComfyUI's native MODEL/CLIP clones using ModelPatcher attachments. Unchanged clones now pass batch validation; mismatched hashes, precision and replaced checkpoints remain rejected.
• Read SFT SafeTensors files with uppercase or mixed-case extensions through the correct reader, including on Windows.

第六次20轮复查完成，修复官方克隆来源标记丢失及SFT扩展名大小写识别问题。新增11项回归，236项测试全部通过。

Validation used real ComfyUI CPU clone interfaces, a one-parameter replacement for diffusion weight allocation, small SFT files and task/lock probes. The suite includes the existing micro CUDA RoPE check. Full-model 4K generation was not repeated. Existing model files and valid task identities remain compatible. Restart ComfyUI after updating.

BluePixel T8 v0.1.7

Fix four issues confirmed during a fifth 20-round review:

• Reject invalid ConvRot groups before conversion can create a floating-point file labelled as INT8.
• Check conversion inputs for changes and finalize source hashes before publishing output, preserving existing output on detected changes.
• Validate and decode the same image bytes so batch results retain the correct input SHA.
• Keep JSON progress safe for Windows legacy stdout encodings while preserving Unicode after parsing.

第五次20轮复查完成，修复INT8分组误标、转换来源记录错配、任务图片读取竞态和Windows进度输出编码问题。新增29项回归，225项测试全部通过。

Validation used small CPU tensors, temporary weights, simulated sampling, real task storage and independent process locks. The suite includes the existing micro CUDA RoPE check. Full-model 4K generation was not repeated. Existing model files and valid task identities remain compatible. Restart ComfyUI after updating.

BluePixel T8 v0.1.6

Fix four issues confirmed during a fourth 20-round review:

• Protect planned inputs from internal done, lock, owner and configuration writes, including path and inode aliases.
• Validate model file identities, precision, VAE tiling and runtime settings at the shared execution entry. CLI and graph batches now use the same checks.
• Return exit code 1 when the CLI reports failed tasks, while preserving its JSON summary.
• Finish empty batches before model loading, including with --no-resume.

第四次20轮复查完成，修复内部状态路径覆盖、执行身份校验、CLI失败退出码和空批次多余加载问题。新增22项回归，196项测试通过。

Validation used temporary weights, simulated sampling, real task storage and ComfyUI CPU node interfaces. Full-weight 4K generation was not repeated. Existing model files and valid task identities remain compatible. Restart ComfyUI after updating.

BluePixel T8 v0.1.5

Fix three issues confirmed in a third 20-round review:

• Reject fractional and boolean seeds instead of silently changing their values; preserve existing identities for valid integer-valued seeds.
• Prevent batch metadata and outputs from overwriting planned input images or source JSONL files, including hardlinks, deduplicated image aliases and later tasks' inputs. CLI checks run before model loading.
• Register BluePixel reload factories for all three components. ComfyUI deep clones retain the diffusion model's f64 / time conventions, text precision adapters and VAE tiling options.

第三次20轮复查完成，修复种子静默转换、输入路径碰撞和三个组件深克隆配置丢失问题。新增20项回归，174项测试通过。

ComfyUI CPU clone tests used actual ModelPatcher APIs with single-parameter components; full-weight 4K inference and physical multi-GPU execution were not repeated. Existing model files and valid default task identities remain compatible. Restart ComfyUI after updating.

BluePixel T8 v0.1.4

Fix issues confirmed in a second 20-round review:

• Load valid legacy PyTorch SFT checkpoints without requiring ZIP-only mmap support.
• Release task locks even if owner-record cleanup fails, and reject invalid metadata filenames before execution.
• Reserve internal metadata directory names case-insensitively, including Windows aliases such as IMG / img.
• Recover interrupted JSONL tails before appending, preserving completed history.
• Include diffusion precision in batch identity, check the connected loader mode, and normalize equivalent integer step values.
• Reject non-finite reference images and batch PNG outputs instead of writing black pixels; support BF16 image preprocessing and output conversion through equivalent FP32 values.

第二次20轮复查完成，修复SFT兼容、锁释放、日志恢复、精度续跑和图像转换等问题。154项测试通过，新增30项回归。

Use the same diffusion_dtype in TaskPlan as the diffusion Loader's dtype. CLI: --diffusion-dtype default|bf16|fp32. Existing default-mode task identities and the three published model files remain compatible. Restart ComfyUI after updating.

This review used small tensors, temporary files and real ComfyUI CPU interfaces. Full-weight 4K generation was validated during the initial release and was not repeated in this review.

BluePixel T8 v0.1.3

Fix issues confirmed in a second 20-round review:

• Load valid legacy PyTorch SFT checkpoints without requiring ZIP-only mmap support.
• Release task locks even if owner-record cleanup fails, and reject invalid metadata filenames before execution.
• Recover interrupted JSONL tails before appending, preserving completed history.
• Include diffusion precision in batch identity, check the connected loader mode, and normalize equivalent integer step values.
• Reject non-finite reference images and batch PNG outputs instead of writing black pixels; support BF16 image preprocessing and output conversion through equivalent FP32 values.

第二次20轮复查完成，修复SFT兼容、锁释放、日志恢复、精度续跑和图像转换等问题。149项测试通过，新增25项回归。

Use the same diffusion_dtype in TaskPlan as the diffusion Loader's dtype. CLI: --diffusion-dtype default|bf16|fp32. Existing default-mode task identities and the three published model files remain compatible. Restart ComfyUI after updating.

This review used small tensors, temporary files and real ComfyUI CPU interfaces. Full-weight 4K generation was validated during the initial release and was not repeated in this review.

BluePixel T8 v0.1.2

Fix five issues found during a 20-round review with an independent subagent:

• Recover tasks whose completion records contain malformed JSON structures or invalid output paths.
• Prevent concurrent conversions from overwriting an existing checkpoint unless overwrite is explicitly requested.
• Match PEFT's regex suffix and configuration-order rules for LoRA rank / alpha patterns.
• Reject cached components after their checkpoint file is replaced, including after replanning a batch.
• Honor the text node's requested default layers when the input CLIP already carries custom layer options.

根据主 Agent 与子 Agent 的 20 轮复查，修复损坏记录续跑、并发覆盖、PEFT 配置匹配、模型缓存身份和默认文本层选择五类问题。

Compatible with the existing three split model files. Download models from t8star/BluePixel-Comfy. Restart ComfyUI after updating so loaders record the new component identity checks.