we're so back / it's so over (w/ Little Squeak)

@superalesha There are other things you can patch on 3090s.

1) For example you can remove the maximum of simultaneous NVENC encoding
github.com/keylase/nvidia…

2) Or the p2p gpu unlock
github.com/aikitoria/open…

Confirmed that 3090 unlock is successful
with 360watt power wall, I got 75T -> 135T (measured in Pytorch)
Which is insane

I used github.com/Zagitta/gpu-un… with commit a9b0c75

Meet Qwen3.5-9B-uncensored-MNN: a 9B model built for on-device AI. Abliterated and uncensored, it runs on mobile via MNN. No cloud needed, just pure local power. 1,090 downloads and counting!

動画生成MiniMax H3派生のComfyUIワークフロー「MiniMax H3 REMIX」がv8.0に更新。
画像や音声、動画素材を柔軟に組み替え可能な設計。手足の歪みや動きの硬さを抑えるLoRAを統合し、動画やデジタルヒューマン制作を効率化。

#MiniMaxH3 #ComfyUI

URLはリプ⬇️

MOSS Image Edit loras.

Just like how Krea 2 Identity Edit could bring out a image edit model from an image gen model, MOSS Image Edit is here to do the same.

I got to go back and work more with MOSS. It's like Krea 2, but with a much better license.
huggingface.co/chinoll/Moss-I…

This is awesome! MOSS for image editing

huggingface.co/chinoll/Moss-I…

JUST IN: UK police arrest 15 y/o boy, seize his devices, & collect his DNA after he posted photo of a local politician edited to look like a rabbit.

Minimax H3 ☺️Singularity Finetuned V2

latest V2 INT8 builds on V1 with expanded action coverage, stronger prompt adherence, a fix for audio reference conditioning, and enhanced visual style rendering.

👇
huggingface.co/WarmBloodAban/…

ComfyUI just fine tuned MOSS to make some MASSIVE images, 9MP+.

huggingface.co/comfyanonymous…

MiniMax H3で継ぎ目のないループ動画を生成するComfyUI拡張が公開。
クロスフェードを用いず、潜在空間の終端と始端を循環させて直接サンプリング。音声の継ぎ目処理やControlNet適用、ループ対応のフレーム補間にも対応。

#MiniMaxH3 #ComfyUI

URLはリプ⬇️

Fibo 2.0 can 1 shot edit my thumbnail style as well.

pretty solid 4 step editor

huggingface.co/realrebelai/Fi…

Fix MiniMax H3's 3 FATAL FLAWS in 1 Workflow! 🎬⚡
Tired of shiny plastic skin, floating slow-motion, and painfully slow high-res renders?
I built an all-in-one ComfyUI pipeline:
1️⃣ Singularity V2: Native fast-action dynamics (No slow-mo LoRAs needed!) + matte natural skin
2️⃣ mpler: 1-node low-res build ➔ 3D latent upscale ➔ refinement
3️⃣ VOSR 2.0 Polish: 1.0x feature refinement for razor-sharp facial details
Full guide & RunningHub workflows below: 👇
🔗youtu.be/n0qavGLdKNE
#MiniMaxH3 #SingularityV2 #SelfLift #ComfyUI #AIVideo #GenerativeAI

tried out the new CorridorKey 2 with a clip where the background isn't plain, very impressive

Indian Cybersecurity Certificate holders be like

MiniMax-H3の8step蒸留モデル「FastH3-Omni」をComfyUIで検証✨

RTX 5090 / R2V / 768×1344 / 5秒
Base 25step：555秒
Omni：185秒
Omni＋Kitchen：125秒
Omni＋Kitchen＋VSA：115秒
※Base以外はすべて8step

最大約4.83倍高速化！

#MiniMaxH3 #ComfyUI

動画① Base vs Omni＋Kitchen
動画② Base n vs Omni＋Kitchen＋VSA

VSA使用時は鼻周辺に点状のアーティファクトが発生。また、低ステップ化による顔のテカリも相変わらず確認できました。

Introducing statistically-lossless Qwen3.8-27B!

Three models, 23.7–30.5 GB: up to 57% smaller weights than BF16.

98–99% measured EAR, versus 97.71% for native FP8.

EAR measures overlap with BF16’s next-token probabilities. A 99% EAR means 99% average probability overlap, here, measured over the reference’s top-10 tokens after renormalization.

The search optimizes EAR directly, assigning more precision to sensitive projections while keeping activations in BF16. The goal is to find the smallest checkpoint while keeping fidelity as high as possible.

vLLM + Humming • Apache 2.0
GGUF versions coming next week!

Get the models:
huggingface.co/collections/ka…

Based on SLQ (arxiv.org/abs/2605.02404) from ISTA’s DASLab (github.com/IST-DASLab), the group behind GPTQ, GSQ and RCO, with customizations to improve the balance between size and fidelity.

More details soon.

seeded noise reveal dialog 🔊
powered by staggered css masks
check the exploded view ↓

We need more cool voice models!
Ming-3.0 TTS just dropped.

open-weights speech gen for stories, short-video narration, product assistants
- designs and clones voices
-takes director-style instructions
- inline paralinguistic tags across 7 languages
- 0.5B model, Elo 1069
- s at 93%

nomorecoding.github.io/Ming-3.0-tts/

I gotta say,

FastH3 Omni really brought the heat. i have a TINY 8gb vram card with some RAM to back it but ive never been lucky to have decent gen times on anything minimax related until now.

im enjoying every second of this!