NidAll/pruna-image-2.1-comfyui-loras

Pruna Image 2.1 LoRAs for ComfyUI

Unofficial ComfyUI conversion of PrunaAI/Pruna-Qwen-Image-2.1. Improved using Qwen. These are LoRA adapters for Qwen/Qwen-Image-2.1, not a standalone model. This repository is not affiliated with PrunaAI or Qwen.

The original adapters have rank 64 and PEFT alpha 128. Their alpha was stored in safetensors metadata, which ComfyUI's LoRA loader does not use. Each file here adds a scalar .alpha tensor for each of the 224 LoRA targets, preserving the intended alpha/rank scale at LoRA strength 1.0. The LoRA A and B weight tensors are unchanged. The source revision is 113e63b.

Adapter	Workflow	Steps	Sigma schedule (including terminal zero)
8-step LoRA	8-step workflow	8	1, 0.933333333, 0.857142857, 0.769230769, 0.666666667, 0.545454545, 0.4, 0.222222222, 0
5-step LoRA	5-step workflow	5	1, 0.94, 0.857142857, 0.666666667, 0.4, 0

The 8-step adapter is the recommended starting point. The 5-step adapter is faster but has lower visual quality. Use one adapter at a time with its matching schedule.

ComfyUI setup

1. Use a recent ComfyUI version with native Qwen-Image-2.1, ManualSigmas, and SamplerCustom support.
2. Download the Qwen-Image-2.1 base files from Comfy-Org/Qwen-Image-2.1 into:
   • ComfyUI/models/diffusion_models/qwen_image_2.1_int8_convrot.safetensors
   • ComfyUI/models/text_encoders/qwen3vl_8b_int8_convrot.safetensors
   • ComfyUI/models/vae/qwen_image_2.1_vae_bf16.safetensors
3. Put either adapter from this repository in ComfyUI/models/loras/.
4. Download and drag the matching workflow JSON into ComfyUI. If your base files use different names or formats, select your installed files in the loader nodes.

The included workflows use only native ComfyUI nodes and are set up for text-to-image at 1024 × 1024. They use Euler, CFG 1.0, LoRA strength 1.0, an empty negative prompt, and the adapter's exact ManualSigmas sequence. ManualSigmas accepts decimal literals; expressions such as 14 / 15 are not evaluated. It also needs the final 0 to finish sampling.

Pruna trained these adapters at approximately 1K resolution for text-to-image and image editing with up to three reference images. The included workflows demonstrate text-to-image; higher resolutions and other settings are outside the stated training coverage. These v0.1 adapters do not yet match the full-step base model's visual quality. See the source model card for the original usage guidance.

Verification and provenance

Both source downloads matched their Hugging Face SHA-256 file hashes. The converted files were checked against ComfyUI's Qwen-Image-2.1 base-model tensor shapes. ComfyUI's LoRA loader recognized all 224 targets per adapter, including 64 patches for the two halves of the fused image-MLP weights, with alpha 128. The workflows were checked for complete links and matching schedules. Full image quality has not been evaluated for these copies.

File	SHA-256
5-step source	021a6228a0fcd217275190b89072416a2f28548e8e18fe049531dc9b77ef89bd
5-step ComfyUI conversion	d7dd5f828dfa1231fcb42fc88fb100ad368a25fc7444351ba625f4f21ec42bc8
8-step source	f0865d68b02511a3a0ed232d9d1aa99cac3a94166f574b38bcbab1a7a297bb15
8-step ComfyUI conversion	5e6d1e9b8a1f7f7d179f32774ded46601f3d6e0ac0e47966b91cbd95ebf655d6

License and attribution

The adapters are distributed under the Qwen RESEARCH LICENSE AGREEMENT, which permits noncommercial research and evaluation use subject to its terms. Read the license before using or redistributing them. The required Qwen attribution and a notice describing this conversion are in NOTICE. Original adapter weights: PrunaAI/Pruna-Qwen-Image-2.1.