Nodes/Eric_Qwen_Edit_Experiments/Eric Qwen Grounded Generate
ComfyUI Node

Eric Qwen Grounded Generate

A ComfyUI node in Eric/QwenImage with 16 inputs and 1 output.

By EricRollei·Created 5 months ago·Updated 3 months ago· 19
Eric Qwen Grounded Generate
  • pipeline
  • ref_image_1
  • ref_image_2
  • ref_image_3
  • ref_image_4
  • image
grounded_prompt
aspect_ratio1:1
output_mp2.00
ref_modevl_only
negative_prompt
steps8
true_cfg_scale4.0
seed0
custom_width1920
custom_height1080
add_generation_prefixfalse
CategoryEric/QwenImage

Inputs (16)

NameTypeDefaultDescription
pipelineQWEN_EDIT_PIPELINE
grounded_promptSTRINGRich prompt from EricGenSearcherNode (or any detailed prompt).
ref_image_1IMAGEPrimary visual reference image.
aspect_ratioCOMBO1:1Output aspect ratio. 16:9 = landscape (~1920×1080 at 2MP) 9:16 = portrait (~1080×1920 at 2MP) 1:1 = square (~1448×1448 at 2MP) custom = use output_width / output_height exactly
output_mpFLOAT2.000.25–16Output megapixels: 0.5 MP → quick test 1.0 MP → ~1024×1024 (1:1) 2.0 MP → ~1448×1448 (1:1) ~1920×1080 (16:9) 4.0 MP → ~2048×2048 (1:1) (slow) Ignored when aspect_ratio=custom.
ref_modeCOMBOvl_onlyHow reference images condition the generation: vl_only (RECOMMENDED for web images): All refs → VL/semantic path only (REF=False). Model understands subject appearance but generates at full quality without pixel-level constraints. Best when references are web search thumbnails. primary_ref: ref_image_1 → VL+REF (pixel level). Others → VL only. Use when ref_image_1 is a high-quality photo. all_ref: All images → VL+REF. Only use with curated high-quality inputs. NOT recommended for web search results.
ref_image_2optIMAGEOptional 2nd reference image.
ref_image_3optIMAGEOptional 3rd reference image.
ref_image_4optIMAGEOptional 4th reference image.
negative_promptoptSTRING
stepsoptINT81–1008 for lightning LoRA, 50 for base model.
true_cfg_scaleoptFLOAT4.01–20
seedoptINT00–18446744073709550000
custom_widthoptINT1920256–8192Used only when aspect_ratio=custom.
custom_heightoptINT1080256–8192Used only when aspect_ratio=custom.
add_generation_prefixoptBOOLEANfalsePrepend a directive telling the model to create a new image. OFF (default): use the grounded prompt as-is. The Qwen-Image model understands image generation from detailed prompts. ON: prepends 'Generate a completely new image...' May fight against the model's edit-focused fine-tuning.

Outputs (1)

NameTypeDescription
imageIMAGE