---
license: other
license_name: per-model
tags:
  - comfyui
  - quantization
  - benchmark
---
# ComfyUI quantization bench: test files

The exact files behind the [ComfyUI Quantization Bench](https://ai-evolution.dev/quant-bench/),
for the [community test kit](https://github.com/LSDJesus/comfyui-quant-bench), which
downloads what your card needs from here. You don't need to fetch anything by hand.

**These are benchmark artifacts, not a model release.** Every model keeps its original
license, listed below. If you want to use a model, get it from its original source.

## What's here

Paths mirror the bench's layout (relative to `/mnt`), so the kit's code runs unchanged.

| Folder                                           | Contents                                                                                                                         |
| ------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------- |
| `flash/downloads/_quant_bench/bakes/<model>/`  | Quantized copies the report tested: `fp8` (fp8 scaled), `int8` (int8 ConvRot), `w4a8` (ConvRot int4 + codebook), `nvfp4` |
| `flash/downloads/_quant_bench/sources/`        | SDXL UNets extracted from their checkpoints, the DMD2 merges, the extracted SDXL VAEs                                            |
| `flash/models/…`, `cold/models/…`          | bf16 originals (the speed baseline) and each family's VAE                                                                        |
| `flash/downloads/_quant_bench/cond/`, `ref/` | Pre-encoded prompts and the RTX PRO 6000 bf16 reference trajectories for the per-step accuracy check                             |

`manifest.json` in the kit repo lists every file with its size and SHA-256.

## Models and licenses

| Report name                            | Files here                                              | Original                               | License                         |
| -------------------------------------- | ------------------------------------------------------- | -------------------------------------- | ------------------------------- |
| Anima base v1.0, Anima turbo v1.0      | bf16 + bakes                                            | _source_                             | Nvidia Open Model License       |
| Z-Image Base, Z-Image Turbo            | bf16 + bakes                                            | Tongyi-MAI                             | _Apache 2.0_                  |
| MOSS-Image base                        | bf16 + bakes                                            | _source_                             | _Apache 2.0_                  |
| MOSS-Image base + DMD2 8-step (merged) | merge + bakes                                           | MOSS-Image base + _DMD2 LoRA source_ | _Apache 2.0_                  |
| Krea 2 Raw, Krea 2 Turbo               | bakes only (bf16 is 26 GB, not needed by ≤24 GB cards) | Krea                                   | KREA 2 COMMUNITY LICENSE        |
| Illustrious-XL v2.0                    | extracted UNet + VAE + bakes                            | OnomaAI                                | Fair AI Public License 1.0-SD   |
| Luminelle v1.0 + DMD2 4-step (merged)  | merge + extracted VAE + bakes                           | Luminelle (Civitai) + DMD2 (tianweiy)  | Fair AI Public License 1.0-SD   |
| Qwen-Image VAE                         | `vae/qwen/`, `vae/anima/`                           | Qwen                                   | Qwen Research License Agreement |
| FLUX.1 VAE (`ae.safetensors`)        | `vae/Flux/`                                           | Black Forest Labs                      | _Apache 2.0_                  |
| FLUX.2 VAE                             | `vae/Flux.2_D/`                                       | Black Forest Labs                      | _Apache 2.0_                  |

If a license doesn't allow redistributing a model or its quantized copies, leave those
files out (the kit's `--families` option lets contributors skip that family) or make the
repo gated so downloaders accept the terms first (the kit passes `HF_TOKEN` through).