WilliamTong (@sonnhost) header
WilliamTong (@sonnhost) profile picture

VFX General

Vietnam
Joined May 2013
Photo by WilliamTong (@sonnhost)
WilliamTong retweeted
We launched @pxld_ai yesterday! I thought I'd share one of DAWN's quieter superpowers: matting. Hair is the classic failure case, with semi-transparent strands and the old background showing through. DAWN keeps every tight curl and loose strand blowing in the wind, then relights them to match the new world.
Introducing DAWN: our foundation model for world harmonisation. GenAI can change the world. Only PXLD does it with: • The performance exactly as shot, not regenerated. • Total 360 degree control of the world, no text prompts needed. • Unlimited clip length and scenes, for film and high-end VFX.
44
69
2
1,052
4,090,982
WilliamTong retweeted
Best Uncensored AI models now : Overall : Qwen3.8-27B Abliterated Small GPU : Gemma 4 12B Abliterated Fastest : Gemma 4 26B-A4B Abliterated Agentic : Qwen3.6-35B-A3B Uncensored Writing : Mistral Small 3.2 Abliterated No surgery : Hermes 4.3-36B Image : FLUX.1-dev Video : Wan 2.2 Ceiling : GLM-5.1 Abliterated (754B, not for home) Where to get them : all LLMs → huggingface.co/huihui-ai (search the name, grab the GGUF) Hermes → huggingface.co/NousResearch FLUX / Wan → ComfyUI + the model's HF page Setup : ollama pull hf.co/<repo>:Q4_K_M ollama run <model> /set nothink that's it. no API key, no account, no card, just few clicks setup more on how to setup below:
Here's 4 steps to run an uncensored AI on your own GPU for $0: I run a local abliterated Gemma 4 for anything I don't want leaving my machine By the end, you'll know how to: - run a model which has no refuses - plug it into Claude Code - do it all for $0, no API key, no account So, let's go step by step: 1. Install Ollama curl -fsSL ollama.com/install.sh | sh you want a recent version, older builds don't speak the Anthropic API and don't support Gemma 4 2. Pull the model ollama pull igorls/gemma-4-12B-it-heretic-GGUF the biggest mistake people make here is grabbing a random "uncensored" tag with no quality numbers this one is abliterated with Heretic. refusals went 99/100 → 0/100 at a KL divergence of 0.028 from the original, so you keep almost all of the base model's capability one thing: it's still Google's Gemma license underneath, fine for personal use, but read the terms before you build a product on it 3. Turn off thinking (the part every other guide skips) Gemma 4 is a hybrid thinking model and the abliteration targets non-thinking mode leave thinking on and the reasoning block eats your token budget and the answer gets cut off ollama run igorls/gemma-4-12B-it-heretic-GGUF /set nothink if you're calling the API instead, add "think": false to the request 4. Plug it into Claude Code (OPTIONAL) ollama launch claude --model igorls/gemma-4-12B-it-heretic-GGUF or manually: export ANTHROPIC_BASE_URL=http://localhost:11434 export ANTHROPIC_AUTH_TOKEN=ollama claude --model igorls/gemma-4-12B-it-heretic-GGUF two gotchas: - bump the context first. Ollama defaults to 4k and Claude Code's system prompt alone is bigger than that, so the model never sees its tool instructions. restart the server with OLLAMA_CONTEXT_LENGTH=32768 ollama serve (setting it in your shell does nothing if Ollama is already running as a service). this needs more VRAM than the 8GB below - Gemma 4 tool calling is still rough, you should expect some failed edits, imo for serious agentic work use qwen3.5 or glm-4.7-flash instead What you need: - 8-12GB VRAM for Q4 (12B) - ~8GB of disk for Q4, ~13GB for Q8 - more of both if you want a big context honest take: - "abliterated" means refusals removed, not quality added. it's still a 12B - fewer refusals ≠ smarter. for real coding work I still use cloud models - you own the weights, your prompts never leave the box. that's the actual win *IMPORTANT: don't use it for anything illegal or harmful. that's on you, not the model now i'm collecting a list of every uncensored model i'm adding to my local stack and which tasks each one replaces for me also planning a big "Uncensored 101" article, still figuring out all the angles save it so you don't lose it
110
464
16
4,944
204,719
WilliamTong retweeted
Avec ce repo GitHub gratuit t'as plus besoin de payer un motion designer 🥵 475 vidéos faites avec Claude Opus 5.5 : des animations, des vidéos explicatives, de la 3D et même des jeux Et chaque démo a son prompt d'origine donc tu prends celle que tu veux et tu la refais pour ton produit 🤖 Une vidéo comme ça une agence te la facture plusieurs centaines d'euros, là tu la sors en quelques minutes et tu la postes tous les jours pour vendre ton app ou ton shop 💰 Je te mets le lien en réponse 👇
7
92
2
1,014
77,899
I might be the most powerful user of this camera motion in 3D and Motion Design history. They can name it after me in the official 3D book.
4
4
78
2,203
WilliamTong retweeted
🚨 Blender artık hobi oldu resmen. Bir dev GitHub’a “cutaway” diye ücretsiz Claude skill’i atmış. Saatlik 3D motion faturalayan stüdyoların hepsi titremeli artık. Bu normalde ücretli kurs olurdu, GitHub’da bedava duruyor. Döngüsü şöyle çalışıyor: 1. ChatGPT’ye bir prompt atıyorsun, referans çıkıyor: beyaz parlak kabuklar, yeşil sadece jantlarda, camın arkasında koyu dişliler 2. Claude o resmi alıyor + “örümcek” diye tek kelime, three.js’te 8 bacak + 32 eklemi komple yazıyor 3. Resmin yanında still pose ediyor, her checklist satırını evet / kısmi / hayır diye puanlıyor, 2. turda senin onayın için duruyor 4. Koyu pikselleri de sayıyor, resmin %80’inden az koyu kalan still turu geçemiyor 5. Film en yakın onaylı build’in hareketlerini birebir kopyalıyor, her 2.6 saniyede yeni parça açıyor 6. Her frame 2.4x render olup 1080p 60 fps’e iniyor, ses aynı script’te -14 LUFS’ta score’lanıyor Input tek ChatGPT resmiydi. Output lens’e zıplıyor. Ücretsiz skill + tüm breakdown aşağıda 👇 teapawt.org/rektwild/status/211928…
3
108
1
1,287
67,633
WilliamTong retweeted
Wow, now we can do muscle sim with this too
Ever since I saw Pixar's movie #elemental of #Pixar , wanted to make the tool they used, Curvenet Deformer in #Blender and I think I finally got it. This tool will be available in #RigCreator in the future. #fanart #commonsideeffects #b3d #blender #rig #rigging #animation
2
29
698
23,863
WilliamTong retweeted
卧槽,大的真来了!! Hugging Face 上最近冒出一个炸裂级的无审查模型,下载直接被冲爆 DeepSeek-V4-Pro 蒸馏进 Qwen3.5 9B 的完全体,量化后只要 5.6GB 左右,一上线就被疯狂下载 实测结果相当凶残: · 拒答率断崖下跌:100 个刁钻测试提示词,原版 100% 拒答,它只拒了 3 个 · 顶级推理蒸馏:吃了 DeepSeek-V4-Pro 约 25 万条高质量数据,逻辑和数学推理直接拉满 · 几乎零损耗去审查:底层用 Heretic 的 ARA 算法消融限制,把道德锁卸掉,智商基本不掉 · 普通电脑本地就能跑:Q4 量化后 5.6GB 左右,显存要求很低,轻薄本也能带得动 完全开源免费,Apache 2.0 协议,想在本地跑一个听话、推理又强的大模型,建议直接收藏 你本地跑大模型时,被拒答最多的是写代码还是角色扮演? huggingface.co/xtools-at/Dee…
79
303
15
2,473
222,216
WilliamTong retweeted
カメラワークの名前24、保存版です。 AIで動画を作る時に役立ちます。 「いい感じで」「映画のような」ではガチャ。 明確に伝えるには、AIとの共通言語で話す。 ← 基本の12個     映える12個 →
🤖 Made with AI
23
271
26
2,348
180,221
WilliamTong retweeted
发现了一个可以免费薅的提示词站点——OpenNana提示词图库,确实够猛! 这站相当硬核,提示词数量已经堆到19888条。 🔗opennana.com
12
70
1
510
45,771
WilliamTong retweeted
🚨ESTO ES UNA LOCURA Un desarrollador chino acaba de lanzar una herramienta GRATUITA capaz de crear vídeos automáticamente para TikTok, Reels y YouTube Shorts. Se llama MoneyPrinterTurbo. Y ya ha superado las 100.000 estrellas en GitHub. Así funciona: 1. Introduces un tema o una palabra clave 2. La IA genera el guion, la narración y los subtítulos 3. Busca el material visual, edita el vídeo y te entrega el resultado final Todo ocurre en un único flujo de trabajo. Guárdate esta herramienta. Estoy seguro de que puedes sacarle muchísimo partido. Enlace abajo👇
72
406
13
3,173
396,822
WilliamTong retweeted
Kai'Sa voguing FX animation breakdown by Fortiche.
1
205
5
2,756
100,989
UI/UX designers thank me later. 1.⁠ ⁠scrolltide.co - 600+ design Prompts and 200+ components with full source code for 3D and scroll-driven sites. 2.⁠ ⁠scrollshaders.com - 117 animated shaders in 11 styles 3.⁠ ⁠footer.design - better footers 4.⁠ ⁠cta.gallery - CTA inspiration 5.⁠ ⁠styles.refero.design - Real product design styles What did I miss?
23
58
2
472
22,499
新たな地平が見える... 久しぶりにblenderが楽しい
2
21
435
16,248
WilliamTong retweeted
🎯 One click. That’s it. AI Camera Tracking for Blender — no tedious setup, no endless tweaking. Just click and track. Free beta available for free, join our discord server. #Blender3D #b3d #BlenderCommunity #CameraTracking #AITracking #VFX #MotionTracking #BlenderAddon #CGI
7
29
4
292
15,904
WilliamTong retweeted
Week 1 Assignment of our Timing Class. Instead of following rigid templates, our students choose references they want to study. This allows them to get the most out of the curriculum and target their specific goals. #VFX #Animation #Arcane #GameDev
297
4
3,692
93,979
折腾了几天,豁然开朗,感谢大佬的不吝指点!!! 这个流程就是前几天酷炫可控制video的pipeline! 第一步:首先RGBMoCap:普通手机/单目实拍视频切入,AI 直接解算 3D 人体网格(SMPL/SMPL-X),零穿戴设备提取骨骼动画。 第二步:骨骼重定向(Retargeting)与对齐。把真人体型平滑映射到目标角色资产,自动解算修复滑步与穿模,输出干净的 3D/ControlNet 控制层。 第三部:MiniMax H3:作为视频生成基座。H3 的动态连贯性和时序物理规律极强,用来扛住大动态、大视角变化,避免动作抽搐崩坏。 第四步:Character-Swap-LoRA:姿态锁死后交由微调 LoRA 重构外观,通过扩散模型逐帧/时序渲染新角色的、服饰与光影。
RGBMoCap→H3-Character-Swap-LoRA
119
160
8
1,212
84,096
WilliamTong retweeted
MiniMax-H3 Person Remover LoRA Erase a person from a video and rebuild the background behind them with MiniMax H3 You pick a person, SAM 3.1 tracks them through the clip and masks them, and this LoRA regenerates the background in overlapping windows so the person is gone. It is an experimental adapter from Akatz Labs, not an official MiniMax release. You need to supply a clean first frame with the person already edited out, at the same size and framing. Start with a continuous shot of about five seconds at 24 fps and windows of 22 frames. It was tested on an RTX 4090. Full GPU loading went past 24 GB, but offloading got renders through. H3 redraws the whole scene, so areas outside the mask are not locked to their original pixels. The showcase picks good results and is not a benchmark. huggingface.co/akatz-ai/Mini…
4
27
1
176
6,296
WilliamTong retweeted
ということで自作クロマキー かなりいいんじゃない?
7
25
1
486
20,737
いや、マジで RGB MoCap やばいな。 海外の開発者が無料で公開した Windows アプリで、 普通の動画が3Dキャラのアニメになる。Codex に導入を手伝わせる手順書まで付いてる👇 teapawt.org/cutetoxicguy/status/21… ・動画から人の動きを3Dの骨組みに変換 ・足の滑りや浮きを、あとから手で直せる ・FBX などで Blender・Unity・Unreal へ書き出し ・MMD 用の VMD にも書き出せる ・付属の手順書(AGENTS.md)を Codex に読ませれば、インストールや AI モデルの準備を手伝ってもらえる つまり 「AIで動きを取って終わり」じゃない。 そのままゲームや映像に使える形まで仕上げられる。 動画から動きを取るモーションキャプチャは、これまで 「足が床を滑る」で止まりがちだった。 RGB MoCap は違う。 足や手が床に着いている区間をタイムラインで指定し、ずれを直してから書き出せる。 しかも驚くのが アプリ本体が、改変も商用利用もできる MIT ライセンスで公開されていること。 AI モデルは手元の PC で動く。 ただし、 64bit Windows と NVIDIA の GPU、約68GB の空き容量が必要(各モデルの利用条件も個別に確認を)。 モーションキャプチャ=高い機材? もう、そうとは限らない。 手持ちの動画から始められる時代になった。 ゲームや3Dアニメを作る人には、かなり刺さるはず。 2026年、 3Dアニメ作りの前提が変わり始めてる。
10
154
13
1,067
57,132