MiniMax H3 Pika Remix | High + Low VRAM, 8 GB+ v1.0

▶ Video guide: how I made a meme edit with it, one cut at a time: https://www.youtube.com/watch?v=n4yO0oqBWdM

Make a clip from your own pictures, or change a video: swap a person, change one thing, copy its moves, or keep it going. Up to four pictures and one video, each with a job you pick.

Discord · YouTube @ausbossdev

How to use it

1. Load a picture in Picture 1 and pick its job on H3 Prompt Writer: person, place, outfit or thing.

2. Type what happens, one line for each thing. @1 is Picture 1: @1 walks to the window and looks out.

3. Pick the length on Seconds.

4. Pick High VRAM for 24 GB or more or Low VRAM for 8 to 24 GB, and press Run.

Everything else is a switch. Pictures turns on Picture 2 to 4. Video and sound adds your clip, its sound, a voice or a song. Extras has Fast preview and Auto prompt. Finish the clip has FlashVSR for a sharper, 2x bigger clip.

What it can do

• Your own place: Picture 2 as place. @1 sits at the counter in @2

• Two people: Picture 2 as person. @1 and @2 sit side by side on a night bus

• A new outfit: Picture 2 as outfit. @1 walks down a runway wearing @2

• Copy a dance: Video on, as copy its moves. @1 dances on a rooftop at sunset

• Put someone in a video: Video on, as swap the person. Name who goes, then add an @1 looks like line.

• Change one thing in a video: Video on, as change something. He wears @1 in place of his cap

• A longer story: Video on, as keep going. Type what comes next.

For only a person swap, the Person Swap workflow is the smaller page.

What you need

Update ComfyUI (0.38.0 or newer; 0.45 or newer on 8 GB; Auto prompt needs 0.42.0 or newer) and AusBoss nodes (2.25.2 or newer) first.

• minimax_h3_ref2va_pruned_int8_convrot.safetensors → models/diffusion_models · 21.0 GB

• qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors → models/text_encoders · 15.7 GB

• minimax_h3_video_vae_fp16.safetensors → models/vae · 5.2 GB

• minimax_h3_audio_vae_fp32.safetensors → models/vae · 0.6 GB

• pika_h3_flex_ref2va_v1_bf16.safetensors → models/loras · 2.0 GB

• Low VRAM · minimax_h3_ref2va_pruned_w6a8.safetensors → models/diffusion_models · 16.0 GB, in place of the first file

• Low VRAM · minimax_h3_video_vae_int8_convrot.safetensors → models/vae · 2.8 GB, in place of the video VAE

• Auto prompt · qwen3vl_8b_int8_convrot.safetensors → models/text_encoders · 9.35 GB

• FlashVSR · flashvsr_v1.1_fp16.safetensors → models/diffusion_models · 3.5 GB

A 1.6 second swap at 864 x 480 took about half a minute on a 32 GB card with the models loaded. Low VRAM uses the two smaller files and makes a smaller clip.

On 8 GB: pick Low VRAM, keep clips to 10 seconds or less, and have about 32 GB of system RAM. We tested by holding back all but 8 GB of a 5090 with RAM capped at 32 GB (ComfyUI 0.45.2): with four pictures a 5 s clip took 39 s, 8 s 72 s and 10 s 87 s, and 12 s ran out of memory; three pictures and a 10 s video took 150 s; an 8 s clip with FlashVSR on took 150 s.

Tips

• Start with Fast preview on and a fixed seed. A draft that looks right can still miss at full size, so check the full size run before you call it done.

• For a swap, the @1 looks like line does most of the work. Without it, the swap can start late or not happen at all.

• A new person, a new place and a new thing all in one go is too much for a few plain lines. In our test the old person came back halfway. Do one change per run, or paste a full H3 prompt.

• A full H3 prompt of your own: paste all six parts, from subject_definitions: to non_diegetic_music:, and it goes to H3 as it is. Your pictures are <Picture 1> to <Picture 4> in the order they are on, and your video is <Video 1>. With a video, set its job to swap the person so the clip keeps the video's length.

• Making a long clip in pieces? Load the last frame of the piece before as a picture, and say in your prompt that the clip continues from it.

• FlashVSR makes the clip 2x larger and sharper (480 x 768 becomes 960 x 1536). It runs with High or Low VRAM.

• Auto prompt needs ComfyUI 0.42.0 or newer. On older versions it wrote a garbage sound line or crashed ComfyUI in our tests.

• Keep the clip's sound also moves the new person's mouth to that sound. With music in the clip, they sing along.

• Leave Pika at strength 1 and add no Turbo or Lightning LoRA. Pika is already the six-step speed-up.

Powered by MiniMax H3 under the MiniMax H3 Community License. The Pika H3 Flex LoRA is by Pika; the ComfyUI files are Comfy-Org's. The optional finish uses FlashVSR v1.1 by Junhao Zhuang and the OpenImagingLab team, under Apache 2.0.

v1.0 (2026-11-13): first release.

Report an issue · Discord · YouTube @ausbossdev · X @Zanzibased · GitHub