The Chinese changelog entry now links to the published Chinese release
notes (v2.6.1.md) instead of the GitHub release page, and the
'(release notes to be published)' note is removed from both EN and ZH.
Documents the LoRA-type ControlNet investigation: full rank-64 LoRA over
all 60 blocks with no injection machinery, ref-concat double-structure
artifact, kv_cache no-steering, not Nunchaku-specific (bf16 also draws
the inner picture). Recommends a standard ControlNet (Alibaba
Qwen-Image-2512-Fun-Controlnet-Union) for Nunchaku Qwen Image. Release
notes link included (to be published).
Expand the Known Limitations entry for LoRA-type ControlNet with the
full background and test results: full-rank-64 LoRA over all 60 blocks
with no injection machinery (Comfy-Org Qwen-Image-DiffSynth-ControlNets /
DiffSynth In-Context Control Union), ref-concat index (double-structure
with inner canny picture), index_timestep_zero (inner picture becomes a
source copy), isolated kv_cache (no steering at all), and the note that
the double structure is not Nunchaku-specific (bf16 also draws the inner
picture; only the outer rendering differs).
Expand the Known Limitations entry for model-patch type ControlNet with
the full technical background: SVDQW4A4Linear hidden-state amplification
(~29x to_qkv, ~760x MLP, block output mean ~1.4e5), per-channel scale
non-uniformity (up to ~1500x), the ~0.3% residual burying, and why
Z-Image (NextDiT) is unaffected. Nunchaku Qwen Image limitation, not a
ControlNet bug.
- Model-patch type ControlNet (DiffSynth block-wise, control_block) is
not usable: W4A4 hidden-state amplification (~29x to_qkv, ~760x MLP,
block output mean ~1.4e5 vs ~0.3-5 standard) buries the standard-scale
residual; naive rescaling breaks channel structure
- LoRA-type ControlNet (qwen_image_union_diffsynth_lora) is not usable:
ref-concat produces the double-structure (inner-picture) artifact on
Nunchaku; isolated kv_cache injection does not steer generation at all
- Recommended: standard (non-model-patch) ControlNet, e.g. Alibaba's
Qwen-Image-2512-Fun-Controlnet-Union via the normal Apply ControlNet flow
The DiffSynth QI route does not support Nunchaku Qwen Image (quantized
hidden-state scale mismatch). Point users to Alibaba's official
Qwen-Image-2512-Fun-Controlnet-Union ControlNet instead.
English technical guide covering symptom, root cause, files changed, full code, and mechanism for the transformers ImageProcessorKwargs 'not documented' startup noise fix (commit 8d82442).
Wraps transformers.utils.auto_docstring.get_args_doc_from_source at prestartup to backfill high_res_size / merge_size / min_pixels / max_pixels that the transformers 5.15.x ImageProcessorKwargs TypedDict docstrings fail to document (deepseek_vl_hybrid, kimi_k25, paddleocr_vl).
- No site-packages modification; runtime wrapper only.
- Auto-skips when upstream transformers fixes the docstrings (subprocess import probe).
- Same design as the existing CausalLM ModelOutput docstring patch.
Complete walkthrough of the cd3ff69 fix: the error, root cause (mmap page-in
failure on damaged precompiled cache), full modified code, and explanation.
load_file() is mmap-based; a cache file with a readable header but damaged
data region crashes the whole process with a Windows fatal 'page error'
(uncatchable). Read the file as plain bytes and parse with safetensors.torch.load
so the same damage surfaces as OSError -> existing except returns {} -> caller
falls back to re-fuse. Verified: broken cache returns empty dict (no crash),
good cache still loads fully.
ussoewwin