Eric Qwen Grounded Generate

Eric Qwen Grounded Generate is the odd one in the pack: it runs on the edit pipeline (QWEN_EDIT_PIPELINE), but instead of editing an image it generates a new one from a detailed prompt plus reference images. The name is a hint - it's purpose-built for a "search the web for references, then generate a grounded result" workflow, sitting after the author's EricGenSearcherNode (his "gen-searcher" research-to-image pipeline) and before a save node.

You hand it a rich prompt (the grounded_prompt input is literally described as "rich prompt from EricGenSearcherNode"), one required reference image plus up to three more, an aspect ratio, and a target resolution. What comes out is a fresh image that understands what your subject looks like - without being a slavish copy of any single reference.

The input that does the heavy lifting: ref_mode

This is the quality lever, and the author's own tooltip is worth quoting: how reference images condition the generation. There are three options:

• vl_only (default, RECOMMENDED) - all references go through the VL/semantic path only. The model's vision encoder understands what the subject looks like, but the generation isn't pinned by pixel-level latents. This gives the highest quality output and is the right call when your references are web thumbnails or scraped images of uncertain quality.
• primary_ref - ref_image_1 goes through both VL and the pixel (VAE/ref) path. Use this only when ref_image_1 is a high-quality photo you specifically want the result to resemble closely. Feed it a low-res thumbnail and you'll feel the quality cost.
• all_ref - every reference goes VL+ref. Only for curated, high-quality input images. The tooltip's warning is blunt: NOT recommended for web search results.

The pattern to internalize (it's the same dual-path design as the rest of this pack, from the KB and README): semantic understanding = free, pixel anchoring = constraining. Start vl_only, escalate only when identity isn't holding.

Everything else

• grounded_prompt - required, the detailed description. The richer, the better; this is a Qwen2.5-VL text encoder, so write sentences, not tag soup.
• ref_image_1 (required), ref_image_2-4 (optional) - the references. The LAST one is resized to target dimensions and sets the output canvas size.
• aspect_ratio - 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9, 9:21, 5:4, 4:5, or custom.
• output_mp - target megapixels (0.25-16). 2.0 MP ≈ 1448×1448 square or 1920×1080 wide; 0.5 MP for quick tests. Ignored in custom mode, where custom_width/custom_height take over.
• add_generation_prefix - off by default. Flipping it on prepends a directive telling the model to create a new image rather than edit the reference. Keep it off unless you see the model "editing" your references instead of generating.
• steps (8 for lightning LoRA, 50 base), true_cfg_scale, seed, negative_prompt - the shared controls.

Output is image.

The honest catch

This is a "generate from references" node built on a 20B edit model, which means it inherits both the model's quality ceiling and its appetite. On consumer hardware it is slow - this pack's generation nodes are the reason people with 5090s report hour-long runs - so set output_mp low while you dial the prompt. And the author's own testers found it most impressive for illustrative/stylized content; the model leans photorealistic and won't match every aesthetic. Install: ComfyUI Manager search "Eric Qwen Edit", or git clone https://github.com/EricRollei/Eric_Qwen_Edit_Experiments into custom_nodes/. You need the Qwen-Image-Edit model loaded (this node takes the edit pipeline, not the generation one - a classic first-run confusion).