NitroFusion2 for ComfyUI | BF16, INT8 GPTQ & W4A8 GPTQ

作者：Silbrio

NitroFusion2 in one or two steps NitroFusion2 is Dar-Yen Chen and Yi-Zhe Song’s text-to-image distillation of Tongyi-MAI’s Z-Image-Turbo. Start with one step for fast prompt and seed exploration. Try two steps if you see artifacts or want to compare the result. The November 7 update adds calibrated GPTQ weights and companion checkpoints for the first step of the two-step workflow. Choose your version

• BF16 Official · 12.31 GB — the original checkpoint and quality reference. For two steps, add the official de-distillation LoRA (0.38 GB). BF16 weights

• INT8 ConvRot GPTQ · 6.18 GB — the most faithful of PedroMarinhoDev’s tested quantizations against BF16. For the recommended two-step workflow, add its 6.18 GB companion. INT8 main · INT8 step 1

• W4A8 GPTQ · 3.49 GB — the smallest of the three downloads, with a greater trade-off in fine detail. For the recommended two-step workflow, add its 3.49 GB companion. W4A8 main · W4A8 step 1

Sizes above use decimal GB and exclude the shared encoder and VAE (8.38 GB together). Civitai may display smaller numbers because it uses binary units. Budget about 21.1 GB for BF16, 20.7 GB for INT8 or 15.4 GB for W4A8 on disk for the recommended two-step setup, including both shared files. PedroMarinhoDev measured about 0.29 s per sampling step for INT8 and W4A8 versus 0.50 s for BF16 at 1024 × 1024 on an RTX 4090, using the initial conversions with the same kernels as the GPTQ files. This timing excludes text encoding and VAE decoding. W4A8 with both checkpoints peaked at 7.9 GB of VRAM in the author’s test. Memory and speed depend on the GPU and ComfyUI setup. Install once

• Use a recent stock ComfyUI build. ComfyUI 0.45.2 is a starting point. The quantization author tested ComfyUI master fb2b26fda with comfy-kitchen 0.2.41. These workflows use built-in nodes.

• Put the main checkpoint and any matching _step1 checkpoint in ComfyUI/models/diffusion_models/. Keep the original file names.

• Put qwen_3_4b.safetensors (8.04 GB) in ComfyUI/models/text_encoders/. The Load CLIP node uses type lumina2.

• Put ae.safetensors (0.34 GB) in ComfyUI/models/vae/. Reuse these two files if you already have them for Z-Image-Turbo.

• Download and unzip the workflow archive under Optional Files for your version. For W4A8, choose the file ending in -r2.zip. Drag the matching one-step or two-step JSON from its workflows/ folder onto the ComfyUI canvas. The archive includes a README and DOWNLOADS.json with upstream links and SHA-256 hashes. Start at 1024 × 1024, enter a prompt and queue it. If a loader shows a missing file, check the folder, refresh ComfyUI and reselect the file.

One step Open the BF16 workflow or the GPTQ workflow. Use the main checkpoint. The settings are 1 step · Euler · Simple · CFG 1 · shift 3. The W4A8 archive includes a preset with the correct model selected. If you use the upstream GPTQ JSON instead, open the subgraph and select the main checkpoint for your format in Load Diffusion Model. Two steps Open the dedicated two-step JSON. It runs two passes on the same schedule with Euler · Simple · CFG 1 · shift 6. Use the dedicated two-step workflow to set up both passes.

• BF16: Put nf2-dedistill-lora.safetensors in ComfyUI/models/loras/ and open the official two-step workflow. It applies the LoRA at 0.5 on the first pass only, then runs unpatched BF16 on the second.

• INT8 or W4A8: Open the GPTQ two-step workflow. First loader: _step1, which already includes the LoRA. Second loader: the matching main file. The per-version ZIPs have both loaders preset. If you use the upstream INT8 JSON for W4A8, change both loaders inside the subgraph.

• GPTQ with less disk space: Use the two-step LoRA workflow with the main GPTQ checkpoint and the official de-distillation LoRA (0.38 GB) in ComfyUI/models/loras/. Open the subgraph and replace the default diffusion model with your main GPTQ checkpoint. The quantization author measured higher error against BF16 with this route than with the paired checkpoints.

Switch back to one step: reopen the one-step JSON and use the main checkpoint. For a side-by-side comparison, set the first sampler’s seed behavior to fixed and keep the same seed, prompt and resolution. Reproduce the gallery Each version has 20 native 1536 × 2048 images, including five anime scenes. The three galleries use the same 20 prompts and seeds, with the two-step workflow, Euler · Simple · CFG 1 · shift 6. To generate at this size, set Width 1536 and Height 2048 on the workflow. Start smaller if GPU memory is limited. Credits and license Unofficial mirror curated by silbrio. NitroFusion2 by Dar-Yen Chen and Yi-Zhe Song, based on Z-Image-Turbo by Tongyi-MAI. GPTQ conversions and calibrated first-step checkpoints by PedroMarinhoDev. The W4A8 workflows in the ZIP preselect the W4A8 files. Each archive includes the full Apache License 2.0 and both GPTQ archives also carry the upstream NOTICE. Official model · Paper · Quantization source and benchmarks · Official LICENSE · GPTQ LICENSE · GPTQ NOTICE.