✦ Enzino H3 Studio · Full edition 🔒

There are two reasonable shapes for a big model's ComfyUI integration. One is the node-per-job chain, which is what most of the ecosystem does and what this pack's WF edition does: loader, encoder, sampler, decoder, saver, each with its own wires. The other is one enormous node with tabs. That's ✦ Enzino H3 Studio - every mode, every knob, a player in the corner. It is a lot of node. Some people find that calming; some find it a wall.

It's Full-edition only, so check that first.

Full edition, and what that means day to day

The pack ships Free and Full builds of the same engine. Free is the six-step WF chain; Full ($11.99, Ko-Fi) adds this node, the Simple card, and the production extras - Krea 2 second pass, story mode, motion control, retouch, variations, 4K, social formats. In the free build EnzinoH3Studio still registers, but as a lock: no inputs, and hitting Run stops the queue with a message naming the Ko-Fi page. Your graph doesn't go red, it just refuses to run.

Be clear-eyed about the edition split, too: the engine is identical across all three front ends, so Studio doesn't render better than the WF chain. What it buys you is density - one place where a whole production lives, instead of six nodes and a ✦ wire between them.

The modes are the point

H3 was trained to treat text, images and audio as one context, and that's why the mode list is long for a video node:

• Text - prompt only. Image - image-to-video, with the picture header written for you.
• First + Last - two keyframes, interpolate between them.
• References - up to nine reference pictures cited in the prompt, for character and product consistency rather than a single anchor frame.
• End frame - the Studio-only one: tell it where the shot lands instead of where it starts. Studio also exposes a mid keyframe with a position from 0 to 1, the closest thing here to real blocking.

Around that sit tabs. Speed is the render: Turbo or Full, seed, sampler and scheduler (ComfyUI's full 44-sampler list, defaulting to res_multistep / simple), plus the memory work - step cache, low-VRAM feed-forward, attention backend, sparse attention, tiled decode, prompt cache. Clip is duration, aspect, size and pixel budget. LoRAs is your stack. Story chains up to twelve shots, each starting on the previous one's last frame, sound crossfaded. Finish is the Krea 2 pass, upscaling and interpolation.

Then the Full-only extras with no tab of their own - variations (2–4 seeds per run, then re-render the winner), motion control through H3's Fun ControlNet, and retouch of a box over a time range - all live in the same node.

How it works under all that

It does what a normal H3 graph does, in the same order, with the loading cached. The text encoder runs first, and only when the prompt actually changed - flip the seed and the 32B Qwen3-VL pass is skipped. The video model is fetched afterwards on purpose, because on a 64 GB machine the encoder and the 33B transformer don't both fit in RAM and Windows starts paging. The Turbo path either applies a Turbo LoRA or, if that slot is empty, treats the diffusion file you picked as already distilled, reading the step count off the filename (4-step file → 4 steps, otherwise 8). Full runs 13 steps with the step cache on, 20 without.

Then attention, sparse attention, chunked feed-forward and the EasyCache patch get applied, and the patched model is cached against a key of the settings that shaped it. Change the seed and it's reused; change the LoRA stack and it's rebuilt - which is why your second clip in a session feels so much faster than the first.

Results land in a player inside the node, with continue-from-last-frame, reuse-seed and restore-settings buttons, a history list, and MP4 (H.264 + AAC) saves with the workflow embedded.

Install

ComfyUI Manager → search the pack title, or:

cd ComfyUI/custom_nodes
git clone https://github.com/byenzinoai/ComfyUI-Enzino-H3-Studio
# restart ComfyUI, hard-refresh the browser (Ctrl+F5)

No pip dependencies - pyproject.toml declares an empty dependency list, because everything runs on ComfyUI's native MiniMax H3 nodes. The requirement is a recent ComfyUI that has them.

The weight bill is the pack's ~40 GB: keyframes model (21 GB INT8, or the 16 GB W6A8 the README recommends on a 16 GB card), a REF2VA file for References mode, the Qwen3-VL 32B encoder in NVFP4 (15.7 GB), the two VAEs, optionally a Turbo LoRA. Use the ◆ Models card rather than downloading by hand - it whitelists the URLs, resumes interrupted downloads, and refreshes ComfyUI's file list without a restart. Language lives under Settings → Enzino → H3 Studio.

Where people get burned

• Folders nested wrong. Forgetting the extra models/diffusion_models and models/text_encoders nesting is the classic H3 install failure. The error names a path - believe the path.
• References mode doing nothing. It needs a REF2VA model or the Hybrid FL+REF file. The WF chain's Render node refuses to run with a clear message; here you just get a clip that ignored your references.
• OOM on a 16 GB card. Leave low-VRAM on auto, use W6A8, install SageAttention, and don't stack 4K upscaling onto a long clip in the same run. The pack retries once after an out-of-memory crash, but it can't invent VRAM.
• Licence. The local H3 weights exclude the EU, UK, South Korea and the US from their applicable territory. The node can't fix that.