Skip to main content
seriesmith/kreagen.py (seriesmith refs <episode>) reads the same refs_todo.json build already writes, generates every image asset on an image target (krea2 turbo unless you choose another), and writes each result straight to the path build expects. No copying files around, no renaming.
The numbers below come from one real episode: 18 image assets — 13 plates, 3 sheets, 2 props — plus 4 voice samples, which are skipped here because they aren’t images. Jobs run most-blocking first, so the lead’s sheet (19 shots) and the busiest plate (6) come out before anything that gates a single shot. Existing files are skipped unless --redo, which can be combined with --only.

​
Which graph runs

The image target decides: --target, else each ref’s override, the episode’s choice (the editor), the series config’s refs.target, else krea2. Any target other than krea2 uses its own workflow (seriesmith/targets/image/<id>/), or one Router call for a hosted one (krea_2_router, gpt_image_2, nano_banana_pro, … : builtin:router_image); an image workflow of your own is a target too (builtin:workflow_image, proposed by the inspector). The rest of this section is krea2. --workflow <file> drives a specific ComfyUI workflow. With no flag, kreagen looks for $KREA_WORKFLOW, then krea2_refs_t2i.json as saved in the running ComfyUI, then $COMFYUI_PATH’s workflows folder, then the repo copy (seriesmith/targets/image/krea2/workflow.json) — found means used. --no-workflow forces the built-in graph, whose model files are in seriesmith/targets/image/krea2/target.json. The run header says which one it took. Per run, --unet, --lora and --lora-strength override what the graph says; --lora splices a LoraLoader into a graph that has none. Per image, kreagen sets only the prompt, size, seed, steps, cfg and save prefix — everything else is yours.

​
Sheets are generated as four separate squares

This is the one real design decision. refs_todo.md asks for a single 4096×1024 four-panel strip, and no diffusion model will give you that — a 4:1 canvas four times wider than anything in its training distribution comes back as smeared repetition. So each view is generated on its own at 1024×1024 and seriesmith/mksheet.py stitches them, which is exactly what mksheet’s docstring says it’s for: “Generate the four views however you like… this assembles them into the one layout H3 actually wants.” Four square panels stitch to exactly 4096×1024, which is the size hint build writes in refs_todo. The four views of one character share a seed, which is most of what keeps them on model. Per-view files land in views/<char>/ so you can regenerate one bad angle and re-stitch without redoing the set. The view prompts are built from the series config’s design sentence plus a framing clause, not from the strip prompt in refs_todo.md — same source text, so the sheet and the 27 shot prompts still cannot drift.

​
Three fixes to the base graph

The krea2 target’s workflow (seriesmith/targets/image/krea2/workflow.json, saved in ComfyUI as krea2_refs_t2i.json) is a krea2 turbo text-to-image graph with three changes from the stock one. kreagen drives that file — save your own copy as krea2_refs_t2i.json among ComfyUI’s workflows, or point --workflow at it — and otherwise builds the same shape of graph in API form from the krea2 target (seriesmith/targets/image/krea2/target.json and graph.py). Driving the file is the better habit: the model, LoRA, sampler and scheduler then live on the canvas, where you can see them. 1. 1280×720 → 1344×768. 720 isn’t divisible by 32. H3’s shot loader raises on it rather than letting H3 fail deep in the graph, so every plate generated at 720 would have been unusable. 2. The LoRA’s CLIP output was going nowhere. In the stock graph, CLIPTextEncode took its CLIP straight off CLIPLoader, so a style LoRA’s strength_clip had no effect at all — you got the model side of the LoRA only. Now, when a LoRA is in play, the text encoder reads the LoRA’s CLIP; --no-lora-clip restores the original wiring, and it’s worth generating one plate each way before committing the set. No style LoRA is loaded by default, because the sheets have to match whatever look series.json asks for and a realism LoRA fights a storybook one. Add one per run with --lora <file> --lora-strength 0.7. For photoreal work a Krea2 realism LoRA earns its place; for flat 2D shows, none usually beats any. 3. Prompt box says that kreagen sets it per image (paste a prompt from refs_todo.md to test by hand) instead of holding a prompt left over from another project.

​
Negatives

Krea 2 has no negative prompt field, and turbo runs at cfg: 1.0, where there is no classifier-free guidance branch to apply one to. Both facts point the same way: on this stack a negative prompt does nothing, which is why the shipped graph wires ConditioningZeroOut instead of a second text encode. kreagen therefore leaves the negative side of a workflow exactly as you saved it. A turbo graph keeps its zeroed branch; a graph built around NAG or negpip — the two nodes that restore suppression at cfg 1 — is passed through untouched. Wire either one in ComfyUI and point --workflow at that file. --negative-file exists for the other case: a model that does expect guidance. Above --cfg 1.0 its text is written to the sampler’s negative input, and an encoder is added if the graph has none. At cfg 1.0 it prints a warning and is ignored. To steer a turbo graph without extra nodes, put the exclusions in the positive text — that is, in the series config’s style.look and design sentences, where they also reach every H3 shot prompt. H3 itself takes no negative prompt at all. The render graph guides with BasicGuider at cfg 1, so an episode’s negative.txt never touches a rendered shot; it only ever shaped reference art.

​
Output paths

Each image is a take (seriesmith/refs.py). A ref with no file yet gets its first good take picked, which copies it to the path above (a character’s sheet is stitched once all four views are picked). --redo adds a take without touching the live file; --pick puts it live. seriesmith/kreagen.py expects seriesmith/mksheet.py beside it — keep both in seriesmith/, or pass --mksheet. Every take is a job on seriesmith’s queue (this process’s own), made on --backend (local, cloud, deployment:<target>; default: each target’s own, usually local, a Router target’s router): the graph keeps its SaveImage, and the picture the backend hands back closes the take. The same command does the rest of a ref’s takes: --clear / --discard a pick or a candidate, --edit REF --instruction "..." (with --with for another ref’s picture) on an edit target, and --voices for the voice refs on the audio target (ltx2_voice by default) or --from-take to cut one out of a rendered take.

​
After a run

mksheet --check reports the strip’s short side and what H3’s ref_image_size: max will do to it. Short side must be ≥1024 or max is upscaling from mush.

​
Order to work in

  1. --only sam first. It gates 19 of 27 shots and sets the look everything else is matched against. Iterate on it alone.
  2. --only core_wide next — most complex space, gates 6 shots, and no plate has been proven under this style yet. If it comes back noisy that’s a location-description problem in the series config, not a graph problem.
  3. Then let the rest run; the ordering is already most-blocking first.