RunningHub UNO Sampler

This is the node people install the pack for. Feed it the three wires from RunningHub UNO Loadmodel, a prompt, and optionally one or more reference images - and it returns a generated image that keeps your subject consistent without a LoRA training run and without an IP-Adapter sitting in your graph. If you've ever wanted to drop a product shot or a logo into a new scene over and over, this is the payoff node.

How it works

UNO treats reference images like IP-Adapter treats CLIP embeddings: it injects them into the Flux transformer as extra tokens alongside your text prompt, so the model is inspired by the reference rather than copying it. Here the trick is baked into a LoRA on the Flux attention blocks (the dit_lora.safetensors your Loadmodel node grabbed), which is why this pack can do it with a plain sampler-shaped node and no external machinery.

Before injection, each reference image gets preprocessed: resized so its long side is 512px (or 320px when you pass several), center-cropped to a multiple of 16, and VAE-encoded. That 512px ceiling is worth remembering - the reference is guidance, not a texture map, so don't expect pixel-for-pixel fidelity on small details.

The inputs that matter

The three uno_model / uno_clip / uno_vae inputs come straight from the Loadmodel node. After that, a beginner really only needs to touch a handful:

• ref_images - optional, and it's the whole point of the pack. Wire an image in and it becomes your subject reference. Leave it empty and you just get plain Flux text-to-image. Pass more than one for multi-reference guidance.
• prompt - the text prompt, multiline. This still does real work; UNO blends prompt and reference rather than letting the image override the words.
• num_steps - defaults to 4, which suits the schnell model. If you loaded flux-dev from the Loadmodel node, raise it (20–30) or you'll leave quality on the table.
• guidance - defaults to 4.0, the right ballpark for dev; irrelevant on schnell.
• width / height - default 704, stepping by 16 (the sampler silently snaps to multiples of 16 anyway). Square images give the best results with UNO.
• seed - defaults to 3407. Community wisdom: some seeds just work better with this model, so if a run looks off, bump it before blaming your prompt.

The one cryptic field is pe, with choices d, h, w, o. It's the positional-encoding offset applied to the reference image's tokens - d (default) shifts them diagonally by half the canvas in both dimensions, h and w shift only height or width, o applies no offset. Think of it as "where the reference content lands relative to the generated image." You can leave it on d for weeks and be fine; if you're fighting composition with multi-image references, this is the knob to experiment with.

Output and gotchas

The single output is image_out (IMAGE), and since this is an output node, it's the end of the line in your graph. Two small quirks from the source: it writes an x.png into ComfyUI's working directory every run (harmless, but it'll clutter things), and it drives a real-time progress bar via ComfyUI's ProgressBar.

For install and model setup, see the Loadmodel article - same pack, same steps. The performance reality from people actually running it: roughly 30 seconds per generation on a 4090, excellent for objects, logos, fashion and product shots, noticeably weaker at consistent faces. That last one is the honest limit of UNO's approach - it nails clothing, hairstyle and tattoos far better than it nails a face. Know that going in and you'll have the right tool for the right job.