Jett Huang

@SlipperyGem They invited everyone to eat "Tuna", but the plate was full of nothing but sauce

Tuna 2 🐟 by Meta has arrived! But ... no weights, because God forbid they actually releases anything. They plan on releasing a lobotomised model. This is so weird. Why bother? github.com/facebookresear…

Part of the NH-Nodes suite: A node for compositing layers onto a background. It supports an unlimited number of layers and allows the use of mathematical expressions to automate any layout you desire. Where: Background width = a - Background height = b Layer 1 width = w1 - Layer 1 height = w2 Each row represents the corresponding x;y coordinates for each layer. Row 1 is assigned x1,y1 - Row 2 is x2,y2, and so on. In the process of building custom workflows, you will occasionally encounter needs like this; instead of having to chain multiple "paste" nodes together, you will only need to use a single node.

I am developing a powerful custom nodes for automation in @ComfyUI , orchestration, and adaptation to various scenarios. This node set is being developed based on projects I have worked on where I encountered automation-related challenges. These custom nodes are currently under development and are being continuously refined. Please use and evaluate them so I can improve them to the best of my ability. github link: github.com/jetthuangai/NH…

@ingi_erlingsson @bfl_ml Great, If the showcase is accurate, you've completely solved the drift issues of Flow Matching models.

FP32 version of the Klein 9B image refiner I’ve been working on and it’s getting so subtle it’s heading into “trust me bro” territory - kind of hard to make out in video form Currently navigating @bfl_ml’s licensing terms to figure out how best to release it 🐣

@m_ai_kel @hectorVFX it's not a 3D model. it's a Video.

@hectorVFX Is the mesh detailed enough for 3d printing ?

Reconstrucción de personaje, 360 + animación con una solo foto utilizando Gaussian Splats y ComfyUI

Este tipo acaba de revelar cómo construir webs cinematográficas de 10.000$ en un tutorial de 16 minutos. La combinación es Gemini 3.1 + Seedance 2.0 y el resultado parece producción de agencia de lujo. Gratis. 16 minutos.

@vivoplt Claude Max 5 and Codex Pro 5x, i bought both them.

Developers, you have $20. What are you buying ? -Claude -Codex

@CrazyVibes_1 China not Viet Nam

VIETNAM 🇻🇳 ! !️ | Young couple fights in front of their little daughter. The man begins to hit the woman in the face, pulls her hair and grabs her by the neck. When his little daughter spoke to him, his fury ceased and that's when the lady began to beat him.

@wildmindai Need to test with Less common face

ComfyUI Flux FaceIR Flux-2-klein LoRA and nodes for blind restoration or reference-guided face restoration github.com/cosmicrealm/Co…

@Depressedprsnn @grok

I haven't used Node 2.0 yet, but with the current @ComfyUI and #custom_nodes, we can simplify and make the workflow extremely streamlined and highly automated.

I have created a set of nodes in @ComfyUI for automatically switching between two process flows without using bypasses, ensuring only a single process executes instead of both simultaneously. Previously, we had to use bypass or mute to prevent them, but this led to an excessive number of connections in the #workflow, or too many groups, and made #API implementation more difficult.

I have just completed a highly efficient character swap workflow with @ComfyUI ; through numerous tests, its stability and accuracy have reached near-maximum levels. Along with that, I have also trained a specialized LoRA for this task with #AITookit of @ostrisai . The workflow for both Klein 4B and 9B are responding very well. #ComfyUI #CharacterSwap #Lora #Tranning

@ai_hakase_ I thought this method has been around for a very long time, hasn't it!?

💥LoRA学習の地獄作業、もうやめませんか？💥 FLUX.2にも対応した最新AI技術「VLM」を使えば、 プロ級の学習データセット作りが完全自動化できちゃうって知ってました？🤯 数百枚の画像へのタグ付けが一瞬で終わる 魔法のような方法を⋯ 👇️ このスレッドで解説していきます！

we are waiting for a new image edit model! FireRed-Image. A versatile high-fidelity editor; - preserves text styles; - restores old photos; - multi-image tasks. It seems to be based on Qwen-Image.. github.com/FireRedTeam/Fi…

Introducing Driftin! I trained DDPM and a new method called "Drifting" on the same 38M-param UNet on CIFAR-10 (classical 8xH100 setup). > DDPM: 50 denoising steps, 418ms per image > Drifting: 1 step, 3.26ms per image > 306 FPS on a single 3090. 57x faster. Same network. > Drifting learns to map noise directly to images in one forward pass using drift fields computed from DINOv2 features. No iterative denoising. No distillation. Just one step. > The quality gap is real at this scale -- but DINOv2 features closed a huge chunk of it, and we have only trained with global batch 1024. Drift signal quality scales directly with batch size. This is the opposite of diffusion. > If the quality gap closes with scale, real-time AI video generation is not a dream -- it is an engineering problem. 77 FPS at 512x512 via latent space on a consumer GPU. Compute sponsored by @VoltagePark Code + devlog below.

Custom workflow with 3D Ghost Mannequin Lora: huggingface.co/nhathoangfoto/…