Poe Image Variation

The "more like this" button

PoeImageVariation is the odd one out in the ComfyPoe pack, because it has no prompt box at all. You hand it an image, and it hands back a variation of that image - same subject, same composition, re-rolled through one of the closed image models on Poe's servers. It's the closest thing ComfyUI has to a "more like this" button, and it's perfect for the moment every generator hits: you made something you actually like, and you want a dozen cousins of it without retyping a prompt that took you an hour to dial in.

It's from the same pack as PoeImageGenerate and PoeImageEdit - a thin wrapper over Poe's OpenAI-compatible API at https://api.poe.com/v1 - so all the same ground rules apply: no local model, no VRAM, needs a Poe account with API access.

How it works

The node encodes your image as a base64 PNG, sends it to Poe alongside an instruction, and turns the first image URL in the reply back into a ComfyUI IMAGE tensor. The interesting part is how the variation is controlled. There's no "denoise strength" slider - instead the variation_strength dropdown (subtle / moderate / strong) selects one of three hardcoded English prompt templates baked into the pack. "Subtle" tells the model to keep things nearly identical, "strong" asks for significant creative changes while holding onto the core concept. That's it. The "strength" is a phrasing choice, not a numeric parameter, so treat it as a vibe selector rather than a precise dial.

The inputs that matter

Only two required:

• image - the IMAGE tensor to vary, from LoadImage or any generator.
• api_key - from https://poe.com/api_key.

Everything else is optional:

• variation_strength - subtle / moderate / strong, default moderate. Start at moderate; it's a good middle ground.
• style_prompt - the one text field this node has, and it's worth using. Whatever you type gets appended as Style: <your text> to the variation instruction. "watercolor", "vintage film grain", "studio product shot" - this is how you steer a variation toward a look.
• model - the same 10-model dropdown as the rest of the pack, default GPT-Image-1.
• aspect - default auto; force 1:1 or 16:9 if you need a specific ratio.
• quality - low / medium / high.
• seed - -1 for random; set one to get the same variation again.

Outputs

image - the variation as a normal IMAGE tensor, straight into PreviewImage or SaveImage. raw_response - the full API reply text, for when the model says something unexpected instead of drawing.

Installing

Search ComfyPoe in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/mrf/ComfyPoe.git

Restart ComfyUI and openai + requests install themselves on first load. No model downloads.

Gotchas

• No prompt, but also no real control. If a variation comes out wrong, your only levers are the strength dropdown, the style_prompt, and re-rolling with a new seed. That's the trade for the simplicity.
• "API key is required" - the field is empty; paste your key in.
• "No image URL found in response" - the model returned text, often a refusal. On DALL-E-3 expect this more than anywhere else in the pack; check raw_response for the actual message.
• Strength is per-model flavor. Because the strength is just English instructions, different models interpret "strong" very differently. Don't be surprised when GPT-Image-1 and Ideogram-v2 give you wildly different takes on the same setting.
• Blocking round-trip. Each variation is a synchronous HTTP call; the node waits on Poe.
• Your API key is in the workflow JSON in plain text. Don't share the file with the key inside.

The honest verdict: this node is a toy in the best sense. It won't give you fine-grained control, but for churning out variations of a render you like - especially with a style_prompt steering the look - it's the fastest path in the pack, and it costs you nothing but a Poe API call.