The pack is inspired by ComfyUI-Pixaroma and serves as a companion collection for nodes that I, Dexogen, wanted in my own workflows. It is not limited to one theme: future nodes may cover any part of ComfyUI.

Nodes

AI Prompt Dexogen External

🧬 Dexogen / 💬 Prompt & Text / AI Prompt Dexogen External

A Pixaroma-style prompt node that uses an external OpenAI-compatible endpoint. It loads the endpoint's model list, supports optional API-key authentication, sampling and reasoning controls, text and reference-image inputs, and built-in prompt presets. Danbooru presets use a local WD SwinV2 tagger, downloaded only when it is first needed.

Video Prompt Dexogen External

🧬 Dexogen / 💬 Prompt & Text / Video Prompt Dexogen External

Creates duration-aware video prompts through an external OpenAI-compatible model. It supports text-to-video, a first-frame reference, and target-specific profiles for MiniMax H3, Wan and LTX. Direct mode passes a prepared prompt through without rewriting it.

Video Storyboard Dexogen External

🧬 Dexogen / 🎬 Video / Video Storyboard Dexogen External

Turns a long brief into a validated MiniMax H3 storyboard. It plans the final timeline on H3's 17k+5 frame grid, accounts for the latent overlap between chunks, and asks an external OpenAI-compatible model only for the story content of the continuity bible and sub-second-to-one-second beats. Python owns IDs, timestamps, bindings, and the final schedule. Every generated chunk prompt is compiled in English with the standard H3 visual, soundscape, and music fields. The operational prompt contains stable subject identity and only the current chunk's states and beats; the full future story bible remains available for review without being repeated as executable conditioning in every chunk. Planning is limited to 16 chunks and one bounded repair pass. Missing secondary detail is tolerated; repair is reserved for structural failures or explicit contradictions with the source.

Storyboard Shot Dexogen

🧬 Dexogen / 🎬 Video / Storyboard Shot Dexogen

Selects one storyboard chunk without contacting an endpoint. The built-in text editor can replace that chunk's compiled H3 prompt without calling the model again. The edited DEXOGEN_STORYBOARD output can go directly to H3 Sequence or through another Storyboard Shot node when several chunks need independent tuning. The original first five outputs keep their positions for existing workflows.

H3 Sequence Dexogen

🧬 Dexogen / 🎬 Video / H3 Sequence Dexogen

Samples the storyboard sequentially as one paired MiniMax H3 video/audio latent. Each continuation uses a fresh unmasked window conditioned by the preceding synchronized video/audio latent tail as a native frame-zero H3 guide. The regenerated hidden overlap is discarded and only the new suffix is appended, without an RGB decode/re-encode seam. Audio windows are aligned to absolute 24 fps / 40 Hz boundaries so long chains do not accumulate a one-token drift. Standard ComfyUI SAMPLER and SIGMAS inputs keep quality or acceleration recipes outside the node. Per-chunk safetensors caches support resume and explicit rerun-from behavior.

The planner and sequence node accept single_clip, continuous_scene, and the experimental edited_sequence mode. Edited sequences still use native H3 latent-tail overlap: the regenerated guided prefix supplies motion context and is discarded from the assembled result. An authored cut, match_cut, dissolve, or fade begins at the first unique beat after that prefix. These are in-chain prompted transitions, not independently rendered or composited shots, and their visual quality must be validated on a real GPU before relying on them in a final render.

The sequence input MODEL is intentionally post-LoRA. Apply detail/anatomy LoRAs first, then at most one acceleration LoRA, then H3 model sampling, and connect the result to H3 Sequence. The node never loads, reorders, or guesses LoRAs from filenames.

H3 Sequence requires ComfyUI with native arbitrary-frame MiniMax H3 video/audio guide support from commit e01fb4c or newer. Older versions fail before sampling.

Long H3 workflow

The included h3_long_scene_gpu_validation.json contains both the long-scene path and a deterministic single-shot inspection path. The saved graph connects Video Storyboard → Storyboard Shot → H3 Sequence, so the selected chunk can be reviewed and edited before sampling. For several edits, chain additional Storyboard Shot nodes. For the complete sequence, decode the same H3 Sequence latent with the MiniMax video VAE and audio VAE, combine them at 24 fps, and save the resulting video. Start with a short two-chunk smoke run before queuing the full validation scene on a CUDA system. The included core LoadImage has no bundled filename: select the actual first frame after opening the workflow. Its IMAGE output is already connected to first_frame on both Video Storyboard and H3 Sequence, so the same graph works for single_clip and continuous_scene without a private placeholder or duplicate image loader. The validation canvas is 1408×768, exactly matching an 11:6 first frame such as 2816×1536; both axes are divisible by 128, so the native fit-disabled first-frame resize preserves the aspect ratio. This is 4.76% more pixels than 1344×768. On an RTX 3090, begin with the documented two-chunk smoke before increasing the sequence length.

Installation

Install ComfyUI-Dexogen from ComfyUI Manager and restart ComfyUI. Pixaroma is recommended for the intended combined collection, but Dexogen can also be installed on its own.

Manual installation:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/dexogen/ComfyUI-Dexogen.git
python -m pip install -r ComfyUI-Dexogen/requirements.txt

Restart ComfyUI and search for Dexogen in the node menu.

License

MIT. Portions adapted from ComfyUI-Pixaroma retain their original attribution and license notice.