Krea 2 + MiniMax H3 - 20260807 - For Patreon Supporters (Patreon)
Videos
-
vid2_00001-audio.mp4
Downloads
Content
Hey everyone.
The public post and YouTube video cover MiniMax H3 Turbo LoRA, Spectrum, Sol-Attn, SageAttention, Sigma Shift, and the prompt agent skill. Those free workflows stop at H3 Turbo + LTX upsampler.
https://www.patreon.com/aifuturetech/posts/165937570
These push further: LTX second pass, FILM frame interpolation, then RTX Video Super Resolution. Same Ref2V and text/image-to-video paths.
Also included: a Krea 2 Character Sheet workflow for building multi-view identity refs before you drop them into H3 Ref2V.
----------------------------------------
WHAT YOU NEED
----------------------------------------
MiniMax H3 + Turbo:
models/loras/minimax_h3_turbo_4step_ema_ckpt500_pruned_comfyui.safetensors
H3 diffusion / CLIP / VAEs from:
https://huggingface.co/Comfy-Org/MiniMax-H3
Turbo LoRA downloads:
https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI
https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora
LTX 2.3 upres:
models/loras/ltx2/ltx-2.3-22b-distilled-lora-384.safetensors
LTX 2.3 checkpoint + text encoder + VAE per your existing LTX install
https://huggingface.co/Lightricks/LTX-2.3
Frame interp + RTX:
film_net_fp16.safetensors
NVIDIA RTX Video Super Resolution node support (RTX GPU)
Krea 2 Character Sheet:
models/diffusion_models/Krea-2-Turbo.safetensors (or krea2_turbo_fp8_scaled / krea2_raw_bf16)
models/text_encoders/qwen3vl_4b_fp8_scaled.safetensors
models/vae/qwen_image_vae.safetensors
models/loras/krea2/QuadView_krea2_v1.safetensors
models/loras/krea2/Krea2-realism-V2.safetensors
CharacterSheet LoRA reference: https://huggingface.co/Alissonerdx/CharacterSheet
Custom nodes:
ComfyUI with MiniMax H3 native nodes (MiniMaxH3ReferenceToVideo, MiniMaxH3ImageToVideo, MiniMaxH3SigmaShift)
In case you want to test Spectrum, and SolAttn
ComfyUI-Spectrum-MiniMax-H3 — https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3
ComfyUI-SolAttn_triton — https://github.com/kijai/ComfyUI-SolAttn_triton
ComfyUI-KJNodes — PathchSageAttentionKJ
https://github.com/kijai/ComfyUI-KJNodes
ComfyUI-LTXVideo
https://github.com/Lightricks/ComfyUI-LTXVideo
ComfyUI-VideoHelperSuite
Frame Interpolation (FILM)
RTX Video Super Resolution
https://github.com/Comfy-Org/Nvidia_RTX_Nodes_ComfyUI
ComfyUI-Easy-Use (or equivalent) — EasyCache / cleanGpuUsed / clearCacheAll
Krea 2 edit nodes — Krea2EditModelPatch, Krea2EditGroundedEncode
Prompt agent skill (optional but used in the video):
https://github.com/benjiyaya/Minimax-H3-Prompt-AgentSkill/
Ref2Vid Test Case :
Prompt:
subject_definitions:<Subject 1> is the young man whose appearance comes from <Picture 1> — a man in his late twenties with an athletic build, oval face with a slightly squared jawline and light stubble, dark brown wavy hair swept back and to the side, dark brown almond eyes, straight nose with a rounded tip, medium lips with a gentle smile. He wears a dark navy-blue short-sleeved crew-neck t-shirt with a vintage graphic print on the chest, medium-wash blue denim jeans, and black low-top sneakers with three white stripes.
<Subject 2> is the young woman whose appearance comes from <Picture 2> — a woman in her early twenties with a slender build, oval face with a defined jawline, dark brunette hair pulled back in a low ponytail with flyaways framing her face, light green-hazel almond eyes, straight nose with a slightly upturned tip, medium lips with natural rose color, fair skin with freckles across her cheeks and nose. She wears a dark forest-green pullover hoodie with drawstrings, olive-khaki knee-length cargo pants, and black low-top sneakers with three white stripes.
summary:[reference generation] <Subject 1> and <Subject 2> stand together holding hands on a beach at sunset. <Subject 1> speaks to <Subject 2>, then <Subject 2> reacts with shock and yells.
retention_analysis:<Subject 1> (appears in [Shot 1], [Shot 2]): fully_preserved - face, dark brown wavy hair, stubble, navy t-shirt, and jeans retained from <Picture 1>.
<Subject 2> (appears in [Shot 1], [Shot 2]): fully_preserved - face, freckles, brunette ponytail, green hoodie, and cargo shorts retained from <Picture 2>.
detailed_description:The target video uses a warm cinematic golden-hour style with soft natural sunset lighting and a warm orange and blue color palette.
[Shot 1] A medium two-shot frames <Subject 1> and <Subject 2> standing side by side on a beach at sunset, holding hands. <Subject 1>, the young man in his navy t-shirt and jeans, faces <Subject 2>. <Subject 2>, the young woman in her forest-green hoodie and cargo shorts, looks up at him. Warm golden sunset light falls across both their faces. The ocean and sky glow orange behind them. <Subject 1>, the young man with dark brown wavy hair and stubble, says with a calm, earnest voice (S1): <d>[English] I will go to MIT soon. Will you wait for me?</d> <Subject 2>'s green-hazel eyes widen. Her lips part slightly. The camera holds on this two-shot.
[Shot 2] At 00:02.500, the shot cuts to a close-up of <Subject 2>, the young woman with the brunette ponytail and freckles. Her expression has shifted to shock — eyebrows raised, mouth wide open. She yells with a raised, emotional voice (S2): <d>[English] What?!</d> Her fair skin and freckles catch the warm sunset light. The camera holds on her shocked face.
overall_soundscape:Gentle ocean waves wash onto the sand in the background. Wind rustles softly. The woman's sudden yell cuts through the calm beach ambience.
non_diegetic_music:A soft acoustic guitar plays a gentle warm melody at a slow tempo. It stops abruptly when the woman yells.
Reference Images:
Generate with : Krea2_CharacterSheet
Output:
----------------------------------------
TIPS
----------------------------------------
Start Turbo at 8 steps for general tests. Drop to 4 when you want speed. Push to 12 when the clip needs more polish.
Keep Turbo LoRA strength around 0.5 in these graphs. Raise only if motion looks under-driven.
If Spectrum or Sol-Attn adds pixelation on far shots or tiny props at 480p-class sizes, bypass those two and leave SageAttention + Sigma Shift on. That matches what I settled on in the video for general purpose.
Resolution Selector at 0.9 MP on 16:9 lands near 1344×768 in these files. Lower megapixels while you lock prompt and character, then raise before the LTX+RTX half.
Run Stage 1, check the H3 Turbo preview, clear GPU cache, then run LTX + FILM + RTX. On tighter cards, do not force the full chain in one go.
Use the prompt agent skill for multi-shot H3 prompts. Tag references the official way and keep shot timing explicit. Vague prompts hurt Ref2V and first/last frame adherence on Qwen3-VL.
Build the Krea 2 character sheet first when you need stable outfit and body views. Export the sheet panels, then feed them into the H3 Ref2V paid graph as Picture references.
On Krea 2 Turbo, stay near 10 steps / cfg 1. Move to 8 if composition drifts. Move to 12 if the face needs more detail.
----------------------------------------
COMMON ISSUES
----------------------------------------
Pixelated far shots with Spectrum or Sol-Attn — bypass those patches; keep SageAttention + Sigma Shift + Turbo.
Soft output after Turbo only — expected on base H3 size. Run the LTX + FILM + RTX stages in these paid graphs.
Turbo LoRA filename mismatch — point Load LoRA at the exact .safetensors on disk (pruned vs ckpt500 pruned). Do not assume both free and paid graphs share one filename.
Out of VRAM on LTX or RTX — lower Resolution Selector megapixels, shorten duration, or run stages separately with clearCacheAll between them.
Missing MiniMaxH3SigmaShift / ReferenceToVideo — update ComfyUI to a build with native H3 nodes.
Audio drift after FILM — keep src fps consistent into combines; FILM doubles frames, so the combine rate path must match.
Krea character sheet views collapse or ignore outfit notes — confirm QuadView + Realism LoRAs are enabled, and keep the grounded encode instruction explicit about face / front / side / back.
Krea Turbo looks soft or under-detailed on faces — raise steps from 10 toward 12 before you leave Turbo for the raw model.
----------------------------------------
Download H3, Turbo LoRA, LTX, and Krea 2 weights before you run. Drop questions in the comments.
Thanks for supporting the channel.
— Benji