W4A4 quantization of Krea-2 Turbo (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). +0.95 dB over the class-heuristic recipe at identical size and speed. The shared modulation projector tproj and the text-fusion transformer stay bf16.

Same seed/prompt, BF16 top vs INT4-mixed bottom. Few-step turbo sampling amplifies trajectory divergence (composition shifts) while style, detail and lighting stay at parity — no quantization artifacts.

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.

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): 17.7 dB (min 14.5) — few-step turbo models amplify trajectory divergence; visual inspection shows style/quality parity with composition shifts, no quantization artifacts.

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