MiniMax H3 RefMod · PRISM (Singularity × 10Eros) — Continue + Outfit Lock

Description

Ref2VA workflow for MiniMax H3 with RefMod face files, running the community PRISM merge (Singularity v2 × 10Eros turbo baked in — no extra turbo LoRA). Built for clip continuation: last frame → AddGuide at frame 0, plus an optional separate outfit/bikini reference (ref_image_0 / <Picture 1>) so wardrobe does not drift when the continue frame is far/back-facing. Manual prompt (no LLM). Auto last-frame PNG next to the saved MP4. Optional SageAttention.

RefMod setup: use StableYogi — MiniMax H3 RefMod (same face, no LoRA) Part 1 to create face mods. Put required face / first-frame / outfit images in ComfyUI/input/ before loading them in this graph.

Features

• PRISM Ref2VA UNET — Singularity + 10Eros turbo in one checkpoint (keep 4-step / 8-step turbo LoRA nodes bypassed)

• RefMod — load face/body .safetensors mods (Part 1 files) → Apply into conditioning

• Continue clips — FIRST FRAME → MiniMaxH3AddGuide only (pose/continuity)

• Outfit lock — separate OUTFIT / BIKINI REF → ref_image_0 = <Picture 1> (clear front/wardrobe), independent from the last frame

• Manual prompt — duration/width/height meta + your prompt → ReferenceToVideo

• Resolution + duration — ResolutionSelector + seconds → frames at 24fps (17n+5)

• Last frame PNG — VHS_SelectImages (-1) + SaveImage (same date stem as the video)

• Optional — Power LoRA rack (off by default), Ref Audio (bypassed), SageAttention (KJ)

How to use

1. Install custom nodes (below) and restart ComfyUI.

2. Download models into the folders listed under Models.

3. Open H3RefMod-PRISM-comfyui.json.

4. RefMod faces (StableYogi Part 1) — follow StableYogi — MiniMax H3 RefMod (same face, no LoRA) to create the face .safetensors. Place those files where the loader finds them, and put any source face / reference images the Part‑1 workflow asks for into ComfyUI/input/. Then in this graph’s Face files, pick mod_1 / mod_2… (press R if the list is empty).

5. FIRST FRAME — last frame of the previous clip into ComfyUI/input/ (continuity → AddGuide only).

6. OUTFIT / BIKINI REF — clear front photo of the same outfit/face also in ComfyUI/input/ → ref_image_0. In the prompt, lock wardrobe with <Picture 1>.

7. Write the prompt in Prompt Manual (subject_definitions + scene; use <Subject 1> matching RefMod names).

8. Set duration (seconds) and Resolution Selector.

9. Queue. Do not enable the turbo LoRA bypass nodes (turbo is already in PRISM).

Prompt tip (continue + outfit)

• Opening state = first_frame / AddGuide (do not call the last frame <Picture 1>).

• <Picture 1> = outfit/bikini/face lock from the OUTFIT REF node only.

Defaults (this graph)

Setting

Value

Steps

8

Scheduler

simple

Sampler

euler

Sigma shift

12 / 3

FPS

24

RefMod

strength 1.0 (adjust copies as needed)

Turbo LoRAs

BYPASS

SageAttention

ON (optional speed)

Custom nodes (dependencies)

Install via ComfyUI Manager or git clone into ComfyUI/custom_nodes/.

Pack

Required

Nodes used

Links

ComfyUI (native MiniMax H3)

Yes

MiniMaxH3ReferenceToVideo, MiniMaxH3AddGuide, MiniMaxH3SigmaShift, VAE/CLIP/UNET loaders, CreateVideo, SaveVideo, ResolutionSelector, …

ComfyUI · MiniMax H3 docs

ComfyUI-MiniMaxH3Mod

Yes

MiniMaxH3RefModsLoader, MiniMaxH3RefModApply

GitHub · comfy.icu

ComfyUI-VideoHelperSuite

Yes

VHS_SelectImages, VHS_LoadAudioUpload (optional audio)

GitHub

ComfyUI-KJNodes

Recommended

PathchSageAttentionKJ, SomethingToString

GitHub

rgthree-comfy

Optional

Power Lora Loader (rgthree)

GitHub

ComfyUI-Custom-Scripts

Optional

PlaySound|pysssss

GitHub

Need a current ComfyUI build with native MiniMax H3 support. RefMod does not replace the H3 core nodes.

Manager search keywords

MiniMaxH3Mod · Video Helper Suite · KJNodes · rgthree · Custom Scripts / pythongosssss

Models

Required

Role

Suggested path under ComfyUI/models/

Source

UNET / diffusion (PRISM)

diffusion_models/MiniMax-H3/singularityV2X10eros_prism.safetensors (name as in this workflow — rename or re-pick after download)

https://civarchive.com/models/3019421/singularity-v2-x-10eros-max

Text encoder (CLIP type minimax)

text_encoders/ e.g. qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors (or official H3 Qwen3-VL)

Comfy-Org/MiniMax-H3

Video VAE

vae/minimax_h3_video_vae_fp16.safetensors

Comfy-Org/MiniMax-H3

Audio VAE

vae/minimax_h3_audio_vae_fp32.safetensors

Comfy-Org/MiniMax-H3

RefMod face files

Put the .safetensors mods where the RefModsLoader expects them (see StableYogi pack / your Part‑1 export folder), then pick them in Face files

Create with ComfyUI-MiniMaxH3Mod · guide + pack: StableYogi — MiniMax H3 RefMod (Civitai)

Optional (in graph, bypassed by default)

Role

Example

Notes

4-step turbo LoRA

loras/Minimax-h3/minimax_h3_ref2v_lightx2v_turbo_4step_…

Keep bypassed on PRISM

8-step turbo LoRA

loras/Minimax-h3/minimax_h3_ref2v_turbo_8step_…

Keep bypassed on PRISM

Power LoRA rack

various NSFW/motion LoRAs

All off unless needed

Ref audio

any short .wav/.mp3/.m4a

Node bypassed

Do not stack LightX2V / Comfy turbo LoRA on top of the PRISM UNET.

Related base models (if you A/B)

• Stock Ref2VA: Comfy-Org/MiniMax-H3

• Singularity: WarmBloodAban/Minimax-h3_Singularity

Node map (this workflow)

Step

Node

Purpose

1

UNETLoader

PRISM Ref2VA

2

CLIPLoader

MiniMax text encoder

3–4

VAELoader ×2

Video + audio VAE

—

MiniMaxH3SigmaShift

Motion dial (12/3)

—

Power Lora Loader

Optional LoRAs

6

MiniMaxH3ReferenceToVideo

Prompt, size, length, refs

7

MiniMaxH3RefModsLoader

Face files

8

MiniMaxH3RefModApply

Inject mods into cond

—

MiniMaxH3AddGuide

First-frame lock @ t=0

—

LoadImage (FIRST FRAME)

Continue last frame → AddGuide only

—

LoadImage (OUTFIT REF)

Clear wardrobe → ref_image_0 / <Picture 1>

—

SamplerCustomAdvanced

8-step euler

—

VAEDecode + VAEDecodeAudio → CreateVideo → SaveVideo

Output MP4

—

VHS_SelectImages (−1) → SaveImage

Last-frame PNG

Outputs

• Video: output/video/H3RefMod-PRISM/H3RefMod-PRISM-<date>_<time>_….mp4

• Last frame: same stem .png (copy to input/ for the next continue)

Credits

• MiniMax H3 — MiniMax / Comfy-Org

• PRISM merge — https://civarchive.com/models/3019421/singularity-v2-x-10eros-max

• Singularity — WarmBloodAban

• RefMod nodes — Luisacaotica/ComfyUI-MiniMaxH3Mod

• Graph inspired by StableYogi-style RefMod Part 2 continue layouts

License / NSFW

Follow the licenses of MiniMax H3, PRISM/Singularity/10Eros, and any LoRAs you enable. This graph can generate NSFW content; use responsibly and only with adult subjects.