Story Panel

Character consistency is the eternal problem in image and video work - you nail a look on frame one and lose it on frame two. Story Panel attacks it from the prompt side: give it a reference image and an LLM backend, and it returns three image prompts describing follow-up shots of the same scene - as if a film crew kept rolling from different positions instead of jumping somewhere new.

The system prompt that drives it (shipped in the pack at system_prompts/story_panel.md) is unusually good at this. It starts by extracting the "visual DNA" of the reference - hair, skin tone, outfit, lighting, lens, framing - and locks those as immutable across all three panels. Then it writes three prompts that are follow-ups in time and space: new angles, new framing, but the same person wearing the same thing in the same place. That's the whole game, and it's the thing diffusion models need spelled out.

How it works

The node resizes your image so the long edge is at most 1568px, then re-encodes it as JPEG, dropping quality until it's under the 5MB API limit. It sends that to Claude or OpenAI (whichever your backend specifies) as a vision call with the story-panel system prompt, then parses the JSON array the model returns into three strings.

If your LLM Backend has reasoning enabled, the node honors it - Claude gets extended thinking with the budget from your backend, which noticeably improves the framing choices on tricky shots. Worth it for hero shots, a waste for B-roll.

Inputs and output

• backend - AI_BACKEND from LLM Backend. Note: unlike Image Prompt by Seed Idea, this one calls Anthropic/OpenAI directly, so it needs a real API key, not a webhook.
• image - the reference frame. Only image[0] is used.
• Outputs: panel_1, panel_2, panel_3 (all STRING) - the three follow-up shot prompts, ready to feed into T2V Prompt Builder or an image sampler.

Installing

Same as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/studiodeep-ai/comfyui-studiodeep StudioDeep
pip install -r StudioDeep/requirements.txt

or search "StudioDeep" in ComfyUI Manager. Restart, then set up an LLM Backend with an Anthropic or OpenAI key - the pack's requirements.txt installs both SDKs.

Troubleshooting

• Strict JSON parse errors - the node regex-hunts for [...] in the model output and json.loads it. If a model refuses to return a clean array, you'll get an exception. Sticking with the models in the LLM Backend's list (Sonnet and up) makes this rare.
• Boring or repetitive panels - that's usually the reference image being too uniform, or reasoning being off. Give it a frame with personality and let the model think a little.
• Big images slow things down - the resize-to-1568px step is automatic, but a 4K source still costs encode time. No action needed; just know it's happening.

It pairs naturally with the pack's T2V Prompt Builder - take panel_1, lock the camera direction, and turn the moment into a video shot.