UNO Generate @REDAIGC

REDUNOGenerate is the payoff half of the QijiTec/ComfyUI-RED-UNO pack - the node that takes one photo of your product, toy, or logo and generates it again in a new scene you describe, with no LoRA training and no adapter fiddling. This is ByteDance's UNO "in-context" generation: the reference image goes straight into the model's context window next to the prompt, the same architectural idea behind FLUX Kontext, rather than being injected as embeddings like IP-Adapter. The difference is UNO was trained for this - and its UnoPE position trick stops the model from blending your reference into the target. The result is "inspired by the reference, not copied": your object stays your object, but the scene, lighting, and angle follow the prompt.

How it works

The node takes the uno_model from REDUNOModelLoader, preprocesses each reference image (long side scaled to 512 for a single reference, 320 when you feed several, then center-cropped to a multiple of 16), and hands them to the UNO pipeline alongside your prompt. Then it runs the Flux-based DiT sampler and hands you back an IMAGE.

The pe parameter is the UnoPE bit. It sets how the reference tokens are offset in the model's position-embedding space so they don't collide with the target image: d (default) shifts both axes, h and w shift one, o shifts nothing. Realistically: leave it on d. The community fiddles with everything else first.

The inputs that matter

Only a few you'll actually touch:

• uno_model - wire in from REDUNOModelLoader. Nothing else accepts it.
• prompt - describe the scene, not the object. The reference owns the object; the prompt owns everything else.
• reference_image_1 through reference_image_4 - optional IMAGE inputs. Wire one LoadImage for a single reference, up to four for multi-subject shots. More references, more identity stability - but each gets preprocessed smaller (320 vs 512), so detail per image drops.
• guidance (4.0), num_steps (25), seed - Flux-style knobs. 25 steps at guidance 4 is the sweet spot; the bundled workflow ships those values.
• width / height - snapped to multiples of 16 under the hood. Square is best, which is less a preference and more a consequence of how references are preprocessed.

Output is a plain IMAGE tensor - wire it into SaveImage or PreviewImage like any other generation node. It also writes a PNG into ComfyUI/output as a side effect, so don't be confused when a file appears without a SaveImage in the graph.

What it's good at (and what it isn't)

The community consensus on FLUX UNO from the workflow-era threads holds here: it shines at objects, fashion, and logos - product shots and mockups where the subject is a thing, not a face. ~30 seconds per generation on a 4090 with fp8, and it's free and local. Where people get burned: faces are mid. UNO locks clothing, hairstyle, and tattoos far better than facial features, so if you need character-consistency for a person, reach for an edit model (Kontext, Qwen-Image-Edit) instead of this. Also keep expectations in check on fine detail - references get downscaled to 512 before inference, so tiny textures that survive on paper don't make it through.

Install and wire-up

Same pack as the loader: search "ComfyUI-RED-UNO" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/QijiTec/ComfyUI-RED-UNO

Restart, and gather the three models from the loader article (RED-UNO FT FP8 checkpoint, the bytedance UNO Dit-LoRA, and the Diffusers-format VAE). The repo ships a working example at example_workflows/flux-red-uno.json if you'd rather not build the graph by hand.

Gotchas

• If you see "Missing keys: 236", it's not this node - the loader got the wrong checkpoint. Fix the base model, not the prompt.
• Feeding all four reference slots because you can isn't free: multi-reference runs shrink each reference to 320px, so you trade per-subject detail for subject count. Two good references beat four mediocre ones.
• The pe dropdown invites experimentation, but there's no evidence most users need anything but the default d. Tweak guidance and seed first - some seeds are just better than others with UNO, so re-roll before you rebuild the workflow.