# Qwen-Image-Lightning Guide to Faster ComfyUI Generation > Published 2026-04-02, updated 2026-09-05 ยท https://www.promptzone.com/neha_sullivan/qwen-image-fast-ai-art-tool-for-comfyui-unveiled-og1 --- title: Qwen-Image-Lightning Guide to Faster ComfyUI Generation published: true description: Configure Qwen-Image-Lightning in ComfyUI, match LoRA weights to the base model, and compare faster sampling against image quality. tags: ai,comfyui,imagegeneration --- Qwen-Image-Lightning accelerates Qwen-Image with LightX2V's distilled LoRA adapters and four-step or eight-step sampling. In ComfyUI, load an adapter with its compatible base weights and use the settings from the matching Lightning workflow. LightX2V provides Apache 2.0 adapter downloads and documents which BF16 and FP8 combinations avoid grid artifacts. [Lightning repository][lightrepo] [Model card][light] ## What are the key facts about Qwen-Image-Lightning? | Field | Verified information | | --- | --- | | Developer | LightX2V develops Lightning; Alibaba's Qwen team develops Qwen-Image. [Lightning repository][lightrepo] [Qwen repository][repo] | | Released | The first eight-step V1.0 adapter was released August 8, 2025; the four-step V1.0 followed August 11. [Release history][lightrepo] | | Type | Distilled LoRA adapters for text-to-image diffusion sampling. [Model card][light] | | Size or parameters | The base Qwen-Image model is 20B; a standalone adapter parameter total is not published in the cited card. [Qwen repository][repo] [Model card][light] | | License and access | Apache 2.0 on the Lightning model card, with downloadable adapter weights. [Model card][light] | | Where it runs | Documented Diffusers CUDA example and ComfyUI workflows. [Model card][light] [ComfyUI guide][comfy] | Lightning changes the sampling configuration needed for generation. Download the adapter as an addition to its documented base model, and keep the complete filenames together in your workflow notes. [Model card][light] ## How does Qwen-Image-Lightning reduce sampling steps? The project provides four-step and eight-step generation routes. Its published examples compare these against the original Qwen-Image setup, giving users a way to investigate whether shorter sampling is useful for their images. This is a concrete change in the inference process, not a guarantee about how many seconds an image takes on your computer. [Lightning repository][lightrepo] A suggested use is composition exploration: try different subject arrangements, backgrounds, and short headlines with an accelerated setup. When a candidate looks useful, compare it with your base-model workflow before deciding which output to keep. The project's evaluations include cases where the relative quality ranking changes. [Lightning repository][lightrepo] The authors also describe V2 adapters as reducing oversaturation compared with V1. That is an attributed model-release claim; use your own skin, fabric, and lighting examples to assess whether the change helps the images you make. [Lightning repository][lightrepo] For reproducible comparisons, keep a small prompt collection and record dimensions, seed, base filename, adapter filename, and settings. Review the same content across candidate configurations. A quicker preview is useful only when it retains the details your project needs. ## What quality and FP8 compatibility limits does Lightning have? The project's evaluation identifies dense or small text and hair-like details as areas where the base model can perform better. It also shows that very complex scenes can fail across all tested configurations. Lower sampling cost therefore should not be treated as proof of equivalent results for every prompt. [Lightning repository][lightrepo] Base-weight compatibility matters. LightX2V documents grid artifacts when its BF16-trained adapters are combined with a directly downcast, unscaled FP8 base. It supplies two remedies: adapters trained for that FP8 base, or a scaled FP8 base compatible with the BF16-trained adapter family. [Compatibility documentation][lightrepo] That compatibility note is more specific than the general ComfyUI tutorial's original Lightning toggle. When assembling an FP8 graph, use a complete pairing from LightX2V's compatibility table and its matching workflow. Do not select weights solely because both filenames contain Qwen-Image. [Lightning repository][lightrepo] [ComfyUI guide][comfy] The ComfyUI guide also cautions against combining its separate distilled checkpoint with the Lightning LoRA. Start with one documented acceleration route and evaluate it before adding other modifications. [ComfyUI guide][comfy] ## How do you set up Qwen-Image-Lightning in ComfyUI? Begin with a working [native Qwen-Image ComfyUI setup](/anika_bose/qwen-image-powerful-ai-art-tool-for-comfyui-lna). Confirm that the base model, Qwen text encoder, and VAE can produce an image before introducing the adapter. This establishes which part of the setup changes when you enable Lightning. For a documented BF16 pairing, open LightX2V's [BF16 four-step workflow][lightworkflow]. It pairs the BF16 base with `Qwen-Image-Lightning-4steps-V2.0.safetensors`. Download the files named by that workflow and place the LoRA in ComfyUI's `models/loras` directory. [Lightning repository][lightrepo] [ComfyUI guide][comfy] Use the sampling settings saved in the chosen graph. The project's command examples pair four-step adapters with four steps and CFG 1.0, and eight-step adapters with eight steps and CFG 1.0. An adapter, its scheduler configuration, and its sampling settings belong together; changing only the step count is not the documented procedure. [Lightning repository][lightrepo] [Model card][light] For a Python CLI route, clone the Lightning repository, follow its environment setup, and download its model repository into a local `Qwen-Image-Lightning` directory. From the cloned code directory, run the documented eight-step V1.0 command; the CLI source defines these flags. [Lightning repository][lightrepo] [CLI source][lightcode] ```bash python generate_with_diffusers.py \ --prompt_list_file examples/prompt_list.txt \ --out_dir test_lora_8_step_results \ --lora_path Qwen-Image-Lightning/Qwen-Image-Lightning-8steps-V1.0.safetensors \ --base_seed 42 --steps 8 --cfg 1.0 ``` The CLI example deliberately names the adapter version its documented command uses. Keep that version visible when comparing it with the separate V2 ComfyUI pairing described above. Refer to the project's Diffusers example for its custom flow-matching scheduler configuration. [Lightning repository][lightrepo] [Model card][light] For broader adapter workflow concepts, see the [LoRA loading and troubleshooting guide](/tara_suzuki/how-to-use-loras-in-comfyui-in-2026-load-stack-and-troubleshoot-235e). Keep a saved base graph alongside the accelerated graph so you can compare lettering and fine detail without rebuilding the setup. ## How does Qwen-Image-Lightning compare with Nunchaku? | Option | Primary change | What to evaluate | | --- | --- | --- | | Original Qwen-Image | The unaccelerated reference generation workflow. [Model card][base] | A baseline for detail and text. | | Qwen-Image-Lightning | Distillation adapters and fewer sampling steps. [Repository][lightrepo] | Quality at the adapter's intended settings. | | Nunchaku Qwen-Image | Specialized quantized weights and runtime; Lightning-fused versions are also provided. [Model card][nun] | Runtime compatibility and memory behavior. | These approaches address different parts of inference. The Nunchaku model card identifies its runtime path and lists weights with Lightning already fused, including their associated adapters and sampling steps. [Model card][nun] ## What should you check before running Qwen-Image-Lightning? ### Does Qwen-Image-Lightning need a base model? The adapters in the Qwen-Image-Lightning model card load alongside the Qwen-Image base pipeline. Nunchaku also publishes quantized checkpoints with named Lightning adapters already fused, which use their own loading instructions. [Lightning card][light] [Nunchaku card][nun] ### Should Qwen-Image-Lightning use four or eight steps? Match Qwen-Image-Lightning's sampling steps to the adapter: four steps for a four-step adapter or eight for an eight-step adapter. The project's examples use CFG 1.0, and its evaluations show that quality varies with the prompt and resolution. [Lightning repository][lightrepo] ### Why does Qwen-Image-Lightning produce grid patterns with FP8? LightX2V documents grid artifacts when BF16-trained Qwen-Image-Lightning adapters are paired with a directly downcast, unscaled FP8 base. Its compatibility table provides an FP8-specific adapter pairing and a scaled-FP8-base pairing to address that mismatch. [Lightning repository][lightrepo] ### How do I measure Qwen-Image-Lightning's speed? Measure both the first and subsequent generations of your complete Qwen-Image-Lightning workflow. ComfyUI reports different timings for those runs on its RTX 4090D test, so keep the hardware, resolution, and model configuration with your timing results. [ComfyUI guide][comfy] ## Sources - [LightX2V Qwen-Image-Lightning model card][light] - [Qwen-Image-Lightning usage, compatibility, and evaluations][lightrepo] - [BF16 base with four-step V2 Lightning workflow][lightworkflow] - [Lightning Diffusers CLI source][lightcode] - [Qwen-Image model card][base] - [Qwen-Image official repository and release history][repo] - [ComfyUI official Qwen-Image workflow guide][comfy] - [Nunchaku Qwen-Image model card][nun] [light]: https://huggingface.co/lightx2v/Qwen-Image-Lightning [lightrepo]: https://github.com/ModelTC/LightX2V-Qwen-Image-Lightning [lightworkflow]: https://github.com/ModelTC/LightX2V-Qwen-Image-Lightning/blob/main/workflows/fp8-comparison/base-bf16-lora-on-bf16.json [lightcode]: https://github.com/ModelTC/LightX2V-Qwen-Image-Lightning/blob/main/generate_with_diffusers.py [base]: https://huggingface.co/Qwen/Qwen-Image [repo]: https://github.com/QwenLM/Qwen-Image [comfy]: https://docs.comfy.org/tutorials/image/qwen/qwen-image [nun]: https://huggingface.co/nunchaku-tech/nunchaku-qwen-image ## Related guides on PromptZone - [ComfyUI 2026: The Complete Guide](/tomas_novak/comfyui-2026-the-complete-guide-to-power-user-ai-image-generation-1g17) - [How to Install and Run SDXL Models in ComfyUI](/jaroslav/how-to-install-and-run-sdxl-models-in-comfyui-a-complete-guide-2nk2) - [How to Use LoRAs in ComfyUI in 2026](/tara_suzuki/how-to-use-loras-in-comfyui-in-2026-load-stack-and-troubleshoot-235e)