Files
mattzh72--articraft/docs/image_conditioned_generation.md
wehub-resource-sync 2c632336aa
CI / Viewer CI (push) Successful in 13m37s
CI / Core CI (push) Has been cancelled
chore: import upstream snapshot with attribution
2026-07-13 12:32:38 +08:00

2.2 KiB

Image-Conditioned Generation

Articraft can use a local reference image alongside a text prompt. The image gives the model visual context, while the prompt should still describe the desired object, articulation, moving parts, and any constraints that matter.

Generate from an Image

Use --image with articraft generate:

uv run articraft generate \
  --image path/to/reference.png \
  "Create an articulated 3D object based on this reference image."

You can combine image input with the normal generation controls:

uv run articraft generate \
  --image path/to/reference.png \
  --max-cost-usd 1.5 \
  "Create an articulated desk fan based on this reference image, with a tilting head and spinning blades."

If you override the model and Articraft cannot infer the provider from the model ID, pass --provider explicitly.

Generate into a Category

Use --image with articraft generate and pass a category slug:

uv run articraft generate \
  --category desk_lamps \
  --image path/to/reference.png \
  "Create an articulated desk lamp based on this reference image."

Edit with a Reference Image

Image input is also supported for copy edits with articraft fork:

uv run articraft fork <record_id> \
  --image path/to/reference.png \
  "Update the object to match the latch shape in the reference image."

The forked record keeps the parent unchanged and stores the edit as a new child record.

Create an Image-Backed Draft

Use articraft draft --image when you want to store a prompt and reference image without running generation yet:

uv run articraft draft \
  --image path/to/reference.png \
  "Create an articulated folding chair based on this reference image."

Supported Image Formats

Supported formats depend on the provider:

Provider Supported formats
openai PNG, JPEG, WebP, GIF
anthropic PNG, JPEG, WebP, GIF
openrouter PNG, JPEG, WebP, GIF
gemini PNG, JPEG, WebP, HEIC, HEIF
deepseek Not supported (no vision capability)

Images must be local files. Gemini image inputs must stay under the inline request limit; other providers reject images larger than 50 MB.