Nodes/ComfyUI-RH-Qwen-Image-2-1-diffusers/RH Qwen-Image 2.1 Diffusers
ComfyUI Node

RH Qwen-Image 2.1 Diffusers

Runs the official diffusers Qwen-Image 2.1 pipeline (Qwen3-VL text encoder + single-stream block-causal transformer + AutoencoderKLQwenImage21). One node covers all four official modes: text-to-image (no image inputs), single-image editing, multi-reference composition (up to 10 images, order matters) and RGBA transparency generation. Weights stream through ComfyUI DynamicVRAM (ModelPatcher + load_models_gpu) per stage.

By RH-RunningHub·Created 4 days ago·Updated 4 days ago· 0
RH Qwen-Image 2.1 Diffusers
  • image_1
  • image_2
  • image_3
  • image_4
  • image_5
  • image_6
  • image_7
  • image_8
  • image_9
  • image_10
  • images
  • alpha
◄prompt►
◄true_cfg_scale1.0►
◄num_inference_steps40►
◄output_resolution1024►
◄width0►
◄height0►
◄seed0►
◄use_kv_cachetrue►
◄negative_prompt►
CategoryRunningHub/Qwen-Image

Inputs (19)

NameTypeDefaultDescription
promptSTRINGPrompt guiding the generation. The pipeline wraps it in the official chat template before the Qwen3-VL encoder.
true_cfg_scaleFLOAT1.01–10Classifier-free guidance scale. Default 1.0 = no guidance (official recommendation). Needs a non-empty negative prompt to activate; doubles the compute per step when enabled.
num_inference_stepsINT404–100Denoising steps. Official recommendation is 40.
output_resolutionINT1024256–2048Target side length. Text-to-image derives a width/height around it; image-conditioned mode derives them from the last condition image's aspect ratio. Rounded to multiples of 32. Use width/height overrides for native-2K ratios (e.g. 2752x1536 16:9).
widthINT00–2752Explicit output width. 0 = derive from output_resolution / condition image aspect ratio. Official 2K 16:9 is 2752.
heightINT00–2752Explicit output height. 0 = derive from output_resolution / condition image aspect ratio. Official 2K 16:9 is 1536.
seedINT00–9223372036854776000Random seed. Same settings + same seed reproduce the image.
use_kv_cacheBOOLEANtrueCache text/condition keys and values after the first step (official default). Disabling it recomputes the whole prefix every step (slower).
image_1optIMAGECondition image slot 1/10. Leave all empty for text-to-image; one image = single-image editing; up to ten = official multi-reference composition (a batched input contributes all its frames in order; later blocks attend to earlier ones, so order matters).
image_2optIMAGECondition image slot 2/10. Leave all empty for text-to-image; one image = single-image editing; up to ten = official multi-reference composition (a batched input contributes all its frames in order; later blocks attend to earlier ones, so order matters).
image_3optIMAGECondition image slot 3/10. Leave all empty for text-to-image; one image = single-image editing; up to ten = official multi-reference composition (a batched input contributes all its frames in order; later blocks attend to earlier ones, so order matters).
image_4optIMAGECondition image slot 4/10. Leave all empty for text-to-image; one image = single-image editing; up to ten = official multi-reference composition (a batched input contributes all its frames in order; later blocks attend to earlier ones, so order matters).
image_5optIMAGECondition image slot 5/10. Leave all empty for text-to-image; one image = single-image editing; up to ten = official multi-reference composition (a batched input contributes all its frames in order; later blocks attend to earlier ones, so order matters).
image_6optIMAGECondition image slot 6/10. Leave all empty for text-to-image; one image = single-image editing; up to ten = official multi-reference composition (a batched input contributes all its frames in order; later blocks attend to earlier ones, so order matters).
image_7optIMAGECondition image slot 7/10. Leave all empty for text-to-image; one image = single-image editing; up to ten = official multi-reference composition (a batched input contributes all its frames in order; later blocks attend to earlier ones, so order matters).
image_8optIMAGECondition image slot 8/10. Leave all empty for text-to-image; one image = single-image editing; up to ten = official multi-reference composition (a batched input contributes all its frames in order; later blocks attend to earlier ones, so order matters).
image_9optIMAGECondition image slot 9/10. Leave all empty for text-to-image; one image = single-image editing; up to ten = official multi-reference composition (a batched input contributes all its frames in order; later blocks attend to earlier ones, so order matters).
image_10optIMAGECondition image slot 10/10. Leave all empty for text-to-image; one image = single-image editing; up to ten = official multi-reference composition (a batched input contributes all its frames in order; later blocks attend to earlier ones, so order matters).
negative_promptoptSTRINGNegative prompt. Only takes effect together with true_cfg_scale > 1 (the model is trained to sample without guidance).

Outputs (2)

NameTypeDescription
imagesIMAGE—
alphaMASK—