Wan 2.1 VACE 1.3B — ConvRot quantized, and the one that proved the threshold is per model

Three quantized builds of Wan 2.1 VACE 1.3B in ComfyUI's native convrot_w4a4 / asym_w4a8_int8 formats. One of them is good. Two of them are here on purpose as examples of what not to ship.

This repo exists because this model refuted a rule the bench had published. Everything below was executed, not inferred, and the prediction was written down before the measurement — and was wrong.

Method, tools and the full log: https://github.com/JoaoZaokk/comfy-quant-bench

---

The three files

Source: 4.01 GiB FP16. The good build is 1.87x lighter; the destroyed one saves another 3.7% and costs you the picture. That is the whole point of publishing all three.

"Effective error" is, per layer, the measured relative error of the format that layer actually received, on the real activations the layer saw during sampling.

What they look like

Six seeds, one prompt, uni_pc / simple, 25 steps, cfg 6.0, ModelSamplingSD3 shift 8, a real negative prompt, vace_strength 0, 33 frames at 480x480. Middle frame of each clip.

Rows top to bottom: FP16 reference · misto005 · misto015 · pure W4A4.

Two seeds at full size, so the degradation is visible without squinting:

misto005 is not "close enough". Across all six seeds it is a sharp watchmaker at a bench, at the same level as the unquantized reference, from a file 1.87x smaller. misto015 keeps the subject and loses everything else to a painterly smear. Pure W4A4 is blocky coloured wreckage in all six.

---

The images on this page were wrong until 2026-09-01, and here is what was wrong with them

Every image previously published here was rendered at cfg 1.0, no shift, no negative prompt. That is the operating point of a distilled model — it is what this bench uses for Z-Image Turbo. Wan 2.1 is not distilled. Without guidance it drifts, and the reference drifted along with everything else: melted faces, smeared arms, washed-out tables, in the FP16 arm that is supposed to be the standard.

Same file, same seeds, only the regime changed:

Top row is what was published. The two rows below are the same unquantized FP16 weights at cfg 6 + shift 8 + negative, and with uni_pc on top of that.

This matters to you as a downloader more than it matters as an apology: if you run these files at cfg 1 with no shift, you will get the top row and blame the quantization. The settings in the section above are not decoration, they are the operating point.

What the correction did not change: the three verdicts. misto005 good, misto015 smeared, pure W4A4 destroyed — all three survived re-rendering in the correct regime across six seeds, and the tolerated / not-tolerated band below was re-judged against the new images rather than the old ones.

One number that did move, and that is worth knowing: misto005's mean divergence from the reference rose from 0.2936 to 0.4750 when the regime was fixed — while the picture got dramatically better. Higher guidance separates trajectories, so that distance measures trajectory, not fidelity. Do not compare divergence numbers across sampler settings, and do not read them as quality.

The owner of this bench caught this by looking at the pictures and asking what the reference was. No automated check on this bench would have. That is worth stating plainly on a page that otherwise argues for measurement.

---

Why this model matters more than its size suggests

The bench had derived a pre-flight rule from two model families: median per-layer W4A4 error above 0.21 breaks, below 0.15 works, in between is correct-but-grainy. Wan 2.1 measures 0.1602 — the middle band, which nothing had ever occupied — and the render is destroyed: no subject, no bench, across three seeds. Not grainy. Gone.

Bracketing with mixed builds puts this model's line between 0.0546 and 0.0793:

The Z-Image row stopped saying not measured on 2026-09-03. Its break was found by moving convrot_groupsize, the only axis left once the model was already 170/170 layers at 4 bits: cg 64 (0.1421) renders fine in 3 of 3 seeds and cg 16 (0.1848) is destroyed in 3 of 3. The prediction written beforehand had the mechanism backwards — it expected a larger rotation group to be worse, and a larger Hadamard rotation actually spreads each outlier across more channels, so it is better. The pre-written control (a smaller group must reduce the error) is what caught that.

That fills the table's last gap, and where it lands is the interesting part: 0.1848 destroys a ~6 B model while 0.1837 is tolerated on a ~13 B one. The bands do not overlap and sit 0.6% apart. Nothing was measured between 0.1421 and 0.1848, so the exact turning point is not a fact; the two ends are.

There is no threshold of the format. There is one per model, and across these three families the tolerated error grows monotonically with model size — now in both columns, not just the tolerated one. The practical consequence is blunt: --promote-error 0.15, a default chosen on a 6B model and carried everywhere since, writes a file here that loads cleanly, dispatches natively, passes every structural check, and renders a smear.

Three points make that a hypothesis, not a law. The next model may break it the way this one broke the last.

Correction, 2026-09-01. An earlier version of this table credited the 0.2163 break to Z-Image, naming capybara_v0.1 as a checkpoint in that family. It is not one. Read from the file, capybara_v0.1 has 1364 tensors and 54 double_blocks — HunyuanVideo 1.5's architecture, not Z-Image's 453 tensors and zero. So that break belongs to the ~13B row, where it agrees with the 0.2147 measured on Hunyuan itself, and Z-Image's upper bound has never been measured: all that is known there is that 0.1241 works. The monotonicity in the tolerated column survives; one cell of evidence moved rows and one cell became honestly empty.

---

The trap that cost four renders, and will cost you the same

A Wan VACE checkpoint sampled without a VACE control node is destroyed by ComfyUI's own defaults — with no error and no warning. At vace_strength 1.0 the unquantized FP16 reference comes out as woven fabric.

Corrected 2026-09-01. This paragraph used to add "in every configuration tried (6 steps/1 frame, 25/33, cfg 6 and cfg 1, with and without ModelSamplingSD3 shift 8)", which read as no setting saves it. That is false and the images above refute it: at vace_strength 0 with cfg 6, shift 8 and a negative prompt, the FP16 reference is excellent. vace_strength 0 is necessary and not sufficient — the sampler settings are the other half, and conflating the two axes is what made this page publish broken references for two weeks.

WAN21_Vace.extra_conds (comfy/model_base.py:1710-1737) fills vace_frames with zeros when no VACE conditioning is supplied, runs each block through process_latent_in — which subtracts the latent format's mean, so zero becomes non-zero — concatenates an all-ones mask, and applies the result at full strength. That is not "no control". It is a constant control at maximum.

If you use these files for plain text-to-video, you must drive vace_strength to 0, or use a real VACE control node. This is a property of the base model's integration, not of the quantization: the unquantized FP16 original fails identically.

---

Speed, and why the sign flips

Same file, same card, only the batch size varied:

The 4-bit kernel carries a per-layer fixed cost that a tiny batch cannot amortise. At a realistic video workload it wins; at one frame it loses. Do not quote a single speed ratio for this format without naming the batch size it was measured at.

---

Verified, not assumed

probe_quant_dispatch.py --forward-only counts, on a real load and a real forward: 300 quantized modules, 12/12 forwards running quantized math, 0 calls to dequantize, linear_dtype int4, implementation resolving to comfy_kitchen.backends.cuda, all quantized weights on cuda:0.

One counting trap worth knowing: loading these prints WARNING: unet unexpected [...comfy_quant] for all 300 layers. It is cosmetic — the tensors are consumed before that check runs and complained about afterwards. It is not a failed load.

Built and verified with:

comfy-kitchen  0.2.31        ComfyUI  c1739380 (0.33.0)
torch          2.13.0+cu130  CUDA     13.0
GPU            RTX 3090 (sm_86)

Native INT4 MMA needs major == 8 (Ampere / Ada). Hopper and Blackwell are routed to an INT8 branch deliberately. The converter hard-refuses to run unless the kernels resolve to comfy-kitchen's CUDA backend, because the eager backend declares the same capabilities and would silently produce numbers describing dequantized math.

---

Format

Standard safetensors, ComfyUI-native, mixed precision in one file — which is the format's own behaviour, not a trick played on it. Per quantized layer, <layer>.weight plus <layer>.weight_scale, and a per-layer entry in __metadata__._quantization_metadata that ComfyUI turns into a <layer>.comfy_quant tensor at load and dispatches on individually. convrot_groupsize 256. Every non-quantized tensor is preserved byte for byte from the source and verified as such. The .quant.json sidecars record full conversion provenance.

The vace_blocks branch is excluded from quantization by design and stays FP16: it only runs when a control input is present, so a calibration without control records nothing for it and the converter would be guessing. Patch/VACE embeddings, norm_q/norm_k, time_embedding, text_embedding and the head are excluded as always.

---

Not covered

One prompt, three seeds (two for the misto005 arm), 480x480, 33 frames, one scheduler, one card. No perceptual metric — "good", "smeared" and "destroyed" are the judgement of someone who looked at them. The checkpoint is FP16, and it is a VACE variant driven as plain T2V, which is not the mode it was trained for: the tolerance measured here may belong to the mode rather than to the model. The calibration sampled one frame; the renders used 33. No SASS. And the monotonicity in model size rests on three points.

---

Credits

• Wan-AI / Alibaba — Wan-AI/Wan2.1-VACE-1.3B, Apache 2.0. This is a derivative of their weights.
• Comfy-Org / comfyanonymous and the ComfyUI contributors — the QuantizedTensor / Layout / MixedPrecisionOps model this format plugs into, and WAN21_Vace itself.
• comfy-kitchen — the ConvRot W4A4 and W4A8 CUDA kernels.
• Quantized by JoaoZaokk with comfy-quant-bench.