Qwen-Image native text-to-image: separate the base, distilled and acceleration routes

File and download boundary: ComfyUI World provides source references and metadata only. Model and LoRA weights, node packages, and workflow files stay with their original providers; this site does not store, mirror, or proxy them. Downloads and access conditions are controlled by each provider.

Purpose and scope

This reference is not a combined download list for Qwen-Image-Edit or every later similarly named model. Select the intended text-to-image template first.

Source-supported context

The official baseline lists qwen_image_fp8_e4m3fn, qwen_2.5_vl_7b_fp8_scaled and qwen_image_vae, with distinct distilled and acceleration-LoRA routes. It does not establish a universally tested combination of every speed-up method. Source 1 Source 2

Before you start

1. The base text-to-image route uses three separate resources and native nodes. Check whether your ComfyUI build includes the example's core nodes; the official guide notes that new nodes may not yet be in the latest stable release.

2. Distilled models need the matching author's sampling/template instructions.

3. Acceleration LoRAs and ControlNets have their own requirements and do not become compatible merely by sharing the Qwen name.

Steps and checks

The sequence below is an editorial procedure based on the cited context, not a diagnosis already confirmed for your environment.

Step 1. Distinguish generation from editing and record the template version and exact files before combining search results.

Step 2. Match diffusion, encoder and VAE to their loaders. Documentation links in the file table still require artifact-revision verification at download time.

Step 3. After the minimal baseline works, enable one acceleration/control route and apply its required sampling settings.

Step 4. For weak effects, structural errors or black output, revert to the baseline and inspect the added branch rather than deleting core encoders.

Completion check

All three components load and the minimal task completes; speed/quality comparisons need equivalent inputs and settings.

Limitations and cautions

File size in GB is not a measured runtime memory requirement. Record author and license separately for base weights, distilled models and LoRAs.

Files named by the source documentation

These are the selected documentation example's declarations, not inspected local files or direct downloads. No weights or checksum results are included; verify the revision for your task.

File	Directory category	Documented source
qwen_image_fp8_e4m3fn.safetensors	diffusion_models	Source
qwen_2.5_vl_7b_fp8_scaled.safetensors	text_encoders	Source
qwen_image_vae.safetensors	vae	Source

Original sources

• Qwen-Image native workflows — original record checked 2026-09-21.
• ComfyUI model troubleshooting — original record checked 2026-09-21.

Localization prepared: 2026-09-22. The source-check dates above were inherited from the existing article; this translation does not claim they were all rechecked today. No GPU run, minimum-memory measurement or runtime guarantee is supplied.

Related troubleshooting and guides

• mat1 and mat2 shapes cannot be multiplied: inspect the model component chain
• CLIP input is invalid: None — trace where the text encoder should come from
• Tune a low-memory workflow one load factor at a time