Image generation
Point LLeMbas at a ComfyUI and a model can draw — against your own workflow, with defaults that suit your hardware.
Set it up
- Go to Admin → Images and enter your ComfyUI address. Test reads its checkpoints, samplers and schedulers.
- Choose a checkpoint and set the defaults: size, steps, CFG, sampler, scheduler, and how many images a run makes.
- Give the drawing tool to the models that should have it under Admin → Models.
Workflows
A workflow is a ComfyUI API-format workflow with placeholders where values go. You declare which node holds what — nothing is guessed from node types.
| Placeholder | Fills |
|---|---|
{{prompt}} | the positive prompt (it can sit inside longer text: "{{prompt}}, masterpiece") |
{{negative}} | the negative prompt |
{{model}} | the checkpoint — ComfyUI’s ckpt_name |
{{width}}, {{height}} | the size |
{{steps}}, {{cfg}}, {{denoise}} | sampling |
{{sampler}}, {{scheduler}} | ComfyUI’s sampler_name and scheduler |
{{seed}} | the seed — random unless the model asks for one; a negative seed means random |
{{batch}} | how many images a run makes — set by the administrator only |
A value that is exactly a placeholder becomes the right type (a number for {{steps}}), so ComfyUI accepts it. The workflow editor shows, beside each placeholder, the value it will have if the model leaves it out.
Defaults
Three layers, the most specific winning: built-in defaults, then this server’s defaults, then what the model asked for. An empty setting means “no opinion”, not zero. The tool’s description tells the model your defaults, so it does not ask for a size your checkpoint draws badly.
Review and retry
Optionally a vision model looks at each picture before you see it — the one chosen in Admin → Images, or else the chat’s own if it has vision. If it rejects one, the tool draws again, up to a limit. The review leans hard towards keeping: a reviewer with taste would spend your GPU four times and end up back at the first picture. The last attempt is always kept, rejected ones are not stored, and anything that goes wrong in review counts as a keep.
Using it
- Ask for a picture in a chat with a model that has the tool, or use
/imageto make it draw now. - A picture is attached to the reply at full quality.
- On a later turn the model cannot see the picture it made, so “make it bluer” draws a new one rather than editing.
- If the GPU runs out of memory, the model is told to retry at a named smaller size or a lighter checkpoint. If you stopped it, it is told not to simply start again.
Sharing one GPU
Preserve VRAM unloads the chat’s own model before drawing, when its connection has an unload address (llama-swap answers one). The language model loads again on the next request. The Images a day quota limits how much one person draws.