LAXMAYDAY/Krea-2-Raw-int4-tensorwise-mixed

Krea-2 Raw — INT4 tensorwise W4A4 (+ConvRot, mixed int8 fallback)

W4A4 quantization of Krea-2 Raw (bf16 source: Comfy-Org/Krea-2): packed signed INT4 weights with group-256 regular-Hadamard ConvRot rotation (arXiv 2512.03673), activations rotated online and dynamically quantized inside the kernel. Calibration-free.

Mixed recipe: 224 block Linears = 128 INT4 + 96 int8_tensorwise fallback, selected by measured per-layer sensitivity: each layer was swapped alone to W4A4 on real sampling inputs, ranked by final-output drift, then chosen by impact-per-byte at a fixed size budget (all wv/wk, 27/28 wo, the most sensitive mlp.down/gate/wq layers). +2.2 dB over the class-heuristic recipe at identical size and speed. The shared modulation projector tproj and the text-fusion transformer stay bf16.

Samples

Same seed/prompt, 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 (NVIDIA L4 / SM 8.9, 1024x1024, 52 steps, cfg 3.5, 3 prompts)

checkpoint	per image
this model (int4 mixed)	207 s
bf16	252 s

Weights: 9.25 GB vs 28 GB bf16 (−67%) — fits a 24 GB card with headroom where bf16 needs weight streaming. PSNR vs bf16 (same seed): 24.3 dB (min 18.6, max 31.0).

Reproduce

comfy-quants export-model-int4-tensorwise \
  --config configs/krea2_int4_tensorwise_mixed.yaml \
  --source krea2_raw_bf16.safetensors \
  --out krea2_raw_int4_tensorwise_mixed.safetensors

Produced by comfy-quants (branch feat/int4-tensorwise). Use with text_encoders/qwen3vl_4b + vae/qwen_image_vae from Comfy-Org/Krea-2.