LAXMAYDAY/Qwen-Image-Edit-2511-int4-tensorwise-mixed

Qwen-Image-Edit-2511 — INT4 tensorwise W4A4 (+ConvRot, mixed int8 fallback)

W4A4 quantization of Qwen-Image-Edit-2511: weights in packed signed INT4 (2/byte) with group-256 regular-Hadamard ConvRot rotation (arXiv 2512.03673), activations rotated online and dynamically quantized to INT4 inside the kernel. Calibration-free.

Mixed-precision quality recipe (validated end-to-end): 480 layers INT4 + 359 sensitive layers (attention value/output projections and the adaLN modulation Linears — critical) stored as stock int8_tensorwise in the same checkpoint. The __index_timestep_zero__ sentinel is included (2511 reference-image mode).

Samples

Same seed/prompt edit samples, BF16 top vs INT4-mixed bottom.

⚠️ Runtime requirement

Needs the int4_tensorwise runtime, which is not in any released ComfyUI / comfy-kitchen yet. In-flight: comfy-kitchen PR #63 (kernels + layout) and the ComfyUI loader branch — to try this checkpoint today, run that ComfyUI branch with the PR's comfy-kitchen branch on PYTHONPATH (or installed). On a stock release it fails with a clear "format not available" error.

Measurements

Quality (L4, 20-step edit sample, mean over 4 prompts, PSNR vs bf16-sentinel):

checkpoint	size	PSNR
int8_tensorwise	20.0 GB	30.05 dB
this model (int4 mixed)	14.9 GB	26.58 dB
bf16	40 GB	—

Speed (NVIDIA L4 / SM 8.9 — native INT4 tensor cores; 20-step image, steady state, same-session comparison, fused ConvRot + rank-one s4 MMA kernel):

path	s/image
this model (int4 mixed)	87.1
int8_tensorwise (CUDA kernels)	93.1
bf16	114

Reproduce

comfy-quants export-model-int4-tensorwise \
  --config configs/qwen_image_edit_2511_int4_tensorwise_mixed.yaml \
  --source /path/to/Qwen-Image-Edit-2511/transformer \
  --out qwen_edit_2511_int4tw_mixed.safetensors

Produced by comfy-quants (branch feat/int4-tensorwise). A higher-quality variant (+MLP down-projections at int8, 17.2 GB / 27.19 dB) is one config edit away — see the config comments.