MiniMax H3 Singularity LoRA · Reference to Video

The interesting thing about this node isn't the model, it's the shape. Every reference group on it - images, videos, audio - is an optional socket that grows another slot the moment you connect one. Nothing tells you what to plug in. You decide how many subjects this shot is casting, and the node grows to fit.

What it is

PopNow_minimax_h3_singularity_lora_reference_to_video renders a new shot featuring subjects you supply, using the community fine-tune of MiniMax H3 rather than the base model. From about 9 credits ($0.09) per second, 3 to 15 seconds, 480p through 1080p, with sound generated in the same pass. It's one node in the PopNow pack, which wraps every PopNow cloud model - so nothing is downloaded, no VRAM is used, and the whole thing works on a laptop.

Reference-to-video is the API-era answer to the consistency problem: instead of training a face LoRA or wiring an IP-Adapter, you hand the model a few photographs and a voice clip and describe a shot with those subjects in it.

How it works, mechanically

The reference fields are built as ComfyUI Autogrow inputs, named reference_image_0, reference_image_1 and so on. Connect one image and a new socket appears; up to nine for images, three for videos, three for audio. On this node all three groups are optional, which means yes - you can run it with no references at all and get a pure text-to-video pass out of the same node. That's not the intended use, but it's a useful way to audition the fine-tune's look.

Everything you attach gets uploaded for you: images as PNG (JPEG fallback above 10 MiB), videos remuxed to MP4, audio converted to 16-bit WAV. Uploads are cached by content hash for 12 hours, so prompt iteration is cheap on bandwidth even though it isn't cheap on credits.

Then you refer to the references by number in the prompt. They condition the generation - they don't become the first frame. If you want the shot to start on your photo, use an image-to-video node instead.

The inputs

• prompt - name references by number and describe the shot they produce, sound included. This is the only thing directing the model, so spend your effort here.
• aspect_ratio - 16:9 default, and this variant adds 9:21 alongside the usual 9:16, 1:1, 4:3, 3:4 and 21:9. Vertical-first work is fine here.
• resolution - 480p, 540p, 768p, 1080p; default 480p. The ladder is the main cost lever.
• duration_seconds - 3 to 15.
• seed - with control-after-generate fixed by default, so re-queuing reuses the result instead of billing you again.
• reference_image_urls (optional) - up to nine; each one adds a per-reference charge.
• reference_video_urls (optional) - up to three, 15 seconds total. Their seconds are billed like output seconds, so a reference clip is real money.
• reference_audio_urls (optional) - up to three clips for a voice or a sound; per-reference charge each.
• loras (optional) - up to three PopNow LoRA nodes on top of the Singularity tuning.

Outputs: video and task id.

Casting kit, not a pile of screenshots

Fewer, better references win. Three photographs - one clear face, one full-body, one costume or product detail - plus five seconds of dry speech beats nine near-duplicates every time, and it's cheaper. Reference clips are worth it when motion matters: a walk cycle, a signature gesture, the way someone turns their head. Keep them short, because you're paying for their seconds.

Install

Manager → search PopNow → Install, or:

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

Restart, paste a key from popnow.ai/api/console/keys into Settings > PopNow, restart again. No Python dependencies are installed - pyproject.toml has an empty list, and the pack rides ComfyUI's own comfy_api.

Practical trouble

PopNow 400 naming a field usually means a reference that didn't fit the model's limits - too many entries, or a video outside the total-seconds budget. PopNow 40201 is credits. The upload ceiling is 90 MiB per video and 10 MiB per image, so shoot-to-edit clips need trimming first. And because references raise the quote, watch the node's price badge after you attach the fifth image - the per-reference charges are small individually and add up quickly.