comfy generate: stale Gemini model enum and incomplete GPT Image 2 output/edit support

[royli001]

comfy generate: stale Gemini image model enum and incomplete GPT Image 2 output/edit support

Describe the bug

comfy generate does not currently expose or fully handle several image models that are available through the current ComfyUI Partner Nodes:

1. The nano-banana adapter uses a fixed Gemini model enum. It accepts the legacy gemini-3-pro-image-preview name, which the server resolves successfully to gemini-3-pro-image, but it rejects the current stable gemini-3-pro-image name and Nano Banana 2 (gemini-3.1-flash-image) before making a request.
2. GPT Image 2 generation is accepted by the server through dalle --model gpt-image-2, but the response contains data[].b64_json. The CLI reports that no image URLs were found and --download does not save the image.
3. The dalle-edit schema does not include an image input, so GPT Image 2 editing cannot be submitted: --image is rejected as an unknown flag.
4. comfy generate refresh currently requests https://api.comfy.org/openapi.yml, which returns HTTP 404, so it cannot refresh the local catalog.

I also checked the current main branch. The Gemini fallback enum, URL-only generic result extraction, and OpenAIImageEditRequest without an image property are still present there.

To Reproduce

Environment:

comfy-cli 1.12.0
Python 3.12
macOS x86_64
Authentication: COMFY_API_KEY environment variable (no browser login)

Inspect the Nano Banana schema:

comfy generate schema nano-banana --json

The model enum is:

gemini-2.5-flash-image
gemini-2.5-flash-image-preview
gemini-3-pro-image-preview

Current stable model names are rejected locally:

comfy generate nano-banana \
  --prompt "A blue ceramic cube on a white background" \
  --model gemini-3-pro-image \
  --json

comfy generate nano-banana \
  --prompt "A blue ceramic cube on a white background" \
  --model gemini-3.1-flash-image \
  --json

Both return:

--model: '<model>' is not one of [...]

However, the legacy Pro name succeeds:

comfy generate nano-banana \
  --prompt "A blue ceramic cube on a white background" \
  --model gemini-3-pro-image-preview \
  --json

The successful server response reports:

{
  "modelVersion": "gemini-3-pro-image"
}

GPT Image 2 generation reaches the server successfully:

comfy generate dalle \
  --model gpt-image-2 \
  --prompt "A red glass sphere on a light gray background" \
  --n 1 \
  --quality low \
  --size 1024x1024 \
  --output_format png \
  --download gpt-image-2.png

The command exits successfully but does not save the image:

No image URLs found in response. Pass --json to inspect.
--download requested but no image URLs found in response.

The JSON response uses data[].b64_json. Trying to force a URL does not work:

comfy generate dalle \
  --model gpt-image-2 \
  --prompt "A green glass sphere on a light gray background" \
  --response_format url \
  --download gpt-image-2.png

The API returns:

API error 400
Unknown parameter: 'response_format'.

GPT Image 2 editing is blocked by the local schema:

comfy generate dalle-edit \
  --model gpt-image-2 \
  --prompt "Replace the circle with a star" \
  --image input.png \
  --download edited.png

Result:

Unknown flag: '--image'.

Finally:

comfy generate refresh

returns:

Failed to fetch https://api.comfy.org/openapi.yml: 404 Not Found

Expected behavior

• nano-banana should accept the current Partner Node model identifiers, including gemini-3-pro-image and gemini-3.1-flash-image, or obtain the enum dynamically from the active catalog.
• --download should decode and save OpenAI data[].b64_json image responses, similar to the existing Gemini inline-image decoder.
• dalle-edit should expose the required image input(s), and optional mask input where supported, so GPT Image 2 editing can be submitted.
• comfy generate refresh should fetch the current authenticated/public OpenAPI catalog from a working endpoint.
• comfy generate list / schema should reflect models currently supported by the official ComfyUI Partner Nodes.

Nice to have

• Terminal output
• Screenshots

Additional context

• Nano Banana Pro generation and image editing both succeeded when using gemini-3-pro-image-preview; the response identified the actual model as gemini-3-pro-image.
• No API keys, account identifiers, generated base64 payloads, or private URLs are included in this report.

[coderabbitai]

🔗 Related PRs

#511 - feat(workflow): agent-first workflow editing — CRDT edit primitives, recipes, offline catalog, cloud run association [open]

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

[dosubot]

Hi @royli001! I'm Dosu and I'm helping the Comfy Org team.

Thanks for the thorough write-up — I was able to confirm each of these against the current codebase:

1. Stale Gemini model enum — The nano-banana adapter's fallback enum is hardcoded to gemini-2.5-flash-image, gemini-2.5-flash-image-preview, and gemini-3-pro-image-preview [1]. The newer gemini-3-pro-image and gemini-3.1-flash-image identifiers are missing. The code is designed so the live OpenAPI spec's enum should take precedence over this fallback [2], but if the spec itself doesn't list them, the stale hardcoded list becomes the effective gate.

2. b64_json not handled — The image extraction path in poll.py only runs _extract_urls(), which walks the JSON response looking for HTTP(S) URLs with media extensions [3]. There is no code path that checks for or decodes data[].b64_json fields, so GPT Image 2 responses are silently dropped.

3. dalle-edit missing image input — The OpenAIImageEditRequest schema declares multipart/form-data but only defines text-level properties (model, prompt, n, quality, size, etc.) with no image or file-upload field, so --image is rejected locally before any request is made.

4. comfy generate refresh 404 — This one was actually addressed in PR #560, which changed the primary endpoint to /openapi (JSON) with a /openapi.yml fallback. If you're on comfy-cli 1.12.0 and still seeing the 404, it's possible this fix hasn't been included in that release yet — upgrading to the latest version may resolve it.

Items 1–3 remain open on main. The b64_json gap (#2) would likely need a decoder similar to the existing Gemini inline-image handler, and the edit schema gap (#3) would need the image (and optionally mask) fields added to OpenAIImageEditRequest.

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.