MiniMax H3 LoRA · Image to Video

The fully-loaded H3 animation node

Opening frame, optional closing frame, up to three LoRAs, seed, and four resolution tiers from 480p to 1080p. If the plain H3 image-to-video node felt stripped down - and it is, deliberately - this is the version with all the controls back.

The LoRA stack is the headline. H3's adapter library is young but real, and being able to lock a look or a subject into a shot is the difference between "generated clip" and "clip that matches the rest of the campaign". The closing frame is the other half: last_image_url gives you an endpoint to resolve on, so you can plan a move from A to B rather than describing A and hoping the camera agrees.

Pricing is per second of output at the tier you pick: about $0.05/s at 480p, $0.075 at 540p, $0.10 at 768p and $0.20 at 1080p. A 5-second 768p clip is around fifty cents; the same length at 1080p is a dollar. The gap between tiers is punitive, so treat 1080p as a finishing move.

Mechanism

Same pack machinery as everything else: media gets uploaded and swapped for a reference only your key can use, the job takes a quote first, submission carries an idempotency key, and the node polls to completion. LoRA files were resolved upstream by PopNow LoRA - either uploaded from ComfyUI/models/loras once (2 GiB limit, cached twelve hours) or imported server-side from a Hugging Face repo, a Civitai page, or a direct link. The result is a VIDEO value heading for Save Video.

Prompting: action, camera, sound - H3 generates audio in the same pass as picture, so describe what it should hear. And keep the LoRA trigger words early; the model doesn't know or care that your adapter wanted to be mentioned, and no upstream node can add it for you.

Inputs

• image_url - the opening frame. The clip keeps its aspect ratio, so there's no aspect_ratio widget here.
• last_image_url (optional) - a closing frame to resolve towards.
• loras - up to three H3 adapters, each with its own scale from -1 to 4.
• resolution - 480p default, then 540p, 768p, 1080p.
• duration_seconds - 3 to 15, whole seconds.
• seed - fixed by default, so an unchanged re-queue reuses the result rather than billing you twice.

Outputs: video and task id.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Then restart. Manager and comfy node install popnow-comfyui are equivalent routes. The pack declares no dependencies and downloads no weights; it's MIT-licensed and uses ComfyUI's own V3 node API, which is why a 33B video model can be installed in two lines.

Key at popnow.ai/api/console/keys → Settings > PopNow, or POPNOW_API_KEY in the environment for headless boxes. Restart once after adding it. Nodes appear under PopNow > Video > MiniMax.

The traps

Family mismatch on the LoRAs. H3 video adapters, not H3 image adapters, not anything else. The upload step doesn't check architecture, so a wrong file uploads fine and then quietly underperforms. The loras tooltip names the family - trust it over the filename.

A closing frame that fights the prompt. If last_image_url shows a bright exterior and your prompt describes a dim interior, you've asked for two different clips. Either match them or drop the closing frame.

1080p out of habit. Four tiers is a trap when the top one costs four times the bottom one. Explore at 480p or 540p, then re-run the winner up a tier. Settings > PopNow > Max credits per run is your seatbelt.

Assuming cancel means refund. It doesn't. The job continues on PopNow and is charged on success - you only stopped watching. Genuine failures (generation_failed) and refusals (content_rejected) are refunded, and PopNow 40201 is simply an empty wallet.