Sam E (@iconsam) header
Sam E (@iconsam) profile picture
New York, NY
Joined August 2006
Photo by Sam E (@iconsam)Photo by Sam E (@iconsam)Photo by Sam E (@iconsam)Photo by Sam E (@iconsam)Photo by Sam E (@iconsam)Photo by Sam E (@iconsam)Photo by Sam E (@iconsam)Photo by Sam E (@iconsam)Photo by Sam E (@iconsam)Photo by Sam E (@iconsam)
Big Recommendation: Lately there's so many new tools and models launching it’s exhausting. But @startvoyager has been a game changer for me. Don't sleep on it. Best way I could explain why was to make this video. + Try it 15% off with promo code: WASS
🤝 Paid partnership
6
5
4
34
1,588
Sam E retweeted
i fine-tuned qwen3-vl 8b to watch a video and tell you how the shot was made. i'm calling it shot reader. camera height, lens, angle, how the camera moves. who's in frame, where they stand, where they walk. turns out that's hard for ai. even opus 5.5 and gemini 4, with a prompt written specifically to reverse-engineer shots, get the camera height and move both right less than half the time. open models out of the box: 5%. shot reader gets 69%. it's 2.4x faster than opus, ~30x cheaper, and small enough to run on your own gpu. why it matters: ai can already write blender. getting a shot right starts with reading it. see a shot → get the plan → rebuild it in 3d, reshoot it, remix it. open weights on hugging face, link below.
72
65
14
986
2,232,195
New stripe press website launches today. Truly a labor of love. Came up with the concept for this website 4 years and in the last year revived and built it with @narrowd and friends. You can read every page of every book, listen to audio book previews, change the lighting, and for the first time, purchase them directly from @stripepress.
118
130
77
1,712
195,568
blockout to video is getting scary good vork to kling 4 + timing it to the beat with voyager
3
3
9
460
blender 5.3 supports gaussian splats. so we built a much faster way to make them. meet vauss — the fastest gaussian splat trainer. • 5.4x faster training than gsplat • about half the gpu memory • better perceptual quality on all 7 comparison scenes works with real video and ai-generated video. real photos → vauss → reusable 3d in blender. the scene below trained in 106 seconds.
24
24
2
272
13,790
Sam E retweeted
Before you burn money generating AI video - it's usually a good idea to do a 3D "previz" which is basically when an agent blocks out a scene for you. Blender is the most complete tool for this - it's free with a GPL license, and Voyager harness recommends using it if you are trying to make an AI film with sufficient complexity. You might have seen that someone on our team released an open-source 3D & vfx tool called Vork which has been getting traction. It is headless right now (no UI for humans), but we're also building a UI plugin for filmmakers where you can control camera location and objects. It's going to be really fast 💨 and f***ing awesome.
12
7
1
44
4,341
I asked Fable 5.2 to "go wild" making a game, and all I got was this Pong clone. And then I played it ... 👇
380
372
284
3,107
416,753
I sent Fable 5.2 on a mission to make a Droste effect using vork from @startvoyager. 15,000 fans. look frame by frame, you'll notice each phone is a viable Droste window. 1080p 59 min render on my M4 MacBook
2
4
7
465
My version, made one-shot by Fable 5.2 in vork. Rendered in 3.5s on M5 Macbook Pro.
🚨 This is made by Fable 5.2. No image model made this. No diffusion, no photos, no textures, no 3D software, no render engine. Every pixel was computed by code Fable 5.2 wrote for this one picture.
4
2
1
49
3,696
Here's an in-depth breakdown of how we used Voyager to create our viral launch video. Voyager handled 700GB+ of cinema camera files (R3D Raw and ProRes 4444), while we steered the entire post production process, from edit, to VFX, to sound, to color grading through the harness. Tools used - Davinci Resolve, Blender, After Effects, Frame.io- all through Voyager.
We just made a 5x cheaper Higgsfield. Voyager gives you AI video models like Seedance 2.5 at half Higgsfield’s price. More importantly, it cuts wasted generations. Users get final outputs in 2–3x fewer API calls. Half the price per generation. A fraction of the generations.
1
3
1
15
2,102
this is awesome. i pushed the other direction a few weeks ago: offline, film-quality honey. honey is one of the hardest things to simulate because it's so thick: it coils, stretches into strings and sticks to things. we tried the open-source options and none handled it well out of the box. so I had claude opus 5.5 agents write a new GPU honey solver from scratch, took an afternoon. it uses the same approach film studios use, 7M particles, and the physics checks out against the textbook math. this 25s pour sims in ~12 min on one H100, about 7-8x faster than the standard published method
Wow, Fable 5.2 is unbelievable at generating photorealistic scenes. "Can you write the most intricate photorealistic simulation of honey in WebGL in the browser" This video is completely physics-based realtime render.
5
6
91
7,324
Rain is a nightmare to film. Rain rigs, waiting on weather, or hours of CG renders just to check a shot. This took 10 minutes in Vork: 150k+ water particles, each refracting the lights in the scene. Total time for @startvoyager to previs it, title comp included: about 30 minutes.
Dear Claude: you are being held back by human-centric graphics APIs. I've built a 3D and VFX engine where agents can write their own shaders, materials, cameras, physics. The core API fits in ~2,500 tokens. Vork is: • 2x more token-efficient vs Blender, Unreal Engine, Houdini • Up to 24x faster rendering times • Rust and Slang based engine, runs on GPU via Metal/Vulkan Below is a simulation of 1 million birds - each with unique trajectories - rendered in 97 sec on one NVIDIA L4 GPU (Blender EEVEE: 31 minutes)
3
3
29
1,858
me: make a space trailer voyager: writes the entire scene in code: ship, hangar, pilot, black hole voyager: casts a voice, composes a score voyager: spins up multiple cloud GPUs, renders 1,291 frames at 1080p voyager: done. under 10 minutes.
3
3
10
753
Sam E retweeted
This is the behind-the-scenes video for Bloom, a 35-second ad for a pen that doesn't exist. Working with world-class product animation directors taught me to treat every product as a character. Both its motion and the camera follow that personality. In Bloom, the pen and cap are actors, so every shot has to have them act! Detail design: - Study real luxury products first. Like Montblanc's continuous body art and engraved gold tip, Bloom has an ink-wash landscape on an ivory body and a hollow rock cap lit from inside. - Design the small parts too. A jade disc hangs on a silk cord, and the tiny mountain engraved on the gold tip became the logo. 3D design: - Plan motion on an exact grey 3D model. A Blender model with exact proportions drove the camera move while design pictures set the look. - Keep the 3D model flat grey with no shine. A shiny white 3D version got copied too closely and came back white without the jade disc or silk cord. Camera and movement: - Hug the product at a steady speed. The camera flies closely along the pen like a drone, after six tries on the grey model to keep the speed even. - Start on extreme close-ups before the full reveal. Early shots get so close the pen looks like a gem, before one final wide shot shows the whole pen standing. Edit: - Cut on movement and check every first and last frame. An early cut chopped a metal ring mid-move, and frame checks caught several shots that had stopped moving. - Give the ending room. After the ink drop cuts to the splash underwater, the ink-wash shot runs longer so the ink turns into a landscape. For details, check out the article.
1
3
7
440
Sam E retweeted
My 53-second fight scene has 57 shots It was all done in one chat, and the only software I opened myself was DaVinci Resolve. I directed it, picked every take and did the grade, while the agent built the Blender previz, wrote the prompts, ran checks and lined up the edit. Takeaways: Set and previz: - Build the room in Blender first. For every new angle, give Nano Banana the Blender frame as the picture to edit so it knows where things go, plus the master shot so it knows the look. The agent had that order the other way round until I told it. Before that, Nano Banana kept moving things around, like stretching the tub all the way to the toilet. - Don't check room angles by eye. The agent laid the Blender edges in red over 6 Nano Banana room pictures. Side by side, 4 of the 6 looked fine, but with the red lines all 6 were off. - When I wanted the room 30% bigger, the agent scaled up a copy of the Blender set just for the reference pictures. The animated previz stayed as it was, so nothing had to be re-blocked. - Keep the previz figures to three blocks: a head, a torso and one block for both legs, with no arms. My first test had full mannequins, and Seedance copied their stiff moves so the actors moved like zombies. Let Blender own the rhythm and camera, and let Seedance do the acting. - Seedance copies whatever is in your previz. The previz door swung 105 degrees in 2 frames, so the door flew open at 1.5x speed. The towel was a rigid bar, so it came out like a board. A white spider-web crack on the Blender mirror came back as neat white lines instead of broken glass. Prompts and continuity: - Keep the shot list to one short line per shot. When the agent wrote out every action and doubled the prompt to over 6,500 characters, the handheld shake died, the camera locked off and cuts vanished. One half-second close-up had five actions written into it, and Seedance stayed on that shot and ate the next two cuts. - Describe what is in the shot, never where it sits on screen. Writing "on the right of frame" gave me a split screen. - Give every prop a labelled picture with its exact size, and call it the same name every time. That was my call, and the lid stopped changing size. Also keep props inside the frame in previz: when the raised lid went above the top edge, Seedance invented a round toilet seat lid instead. - One state per generation. If you mention an injury in the prompt or show it on a reference sheet, the actor has it from frame one, and writing "from 1.6s" does nothing. Pictures also beat text: the prompt said the towel was on the floor, the room pictures had it on the rail, and Seedance put it on the rail. - To keep continuity & look across separate runs, give Seedance the last 10 seconds of finished footage, plus the new Blender previz for the cuts. Fixing a bad shot: - In 7 frames of the lid strike, the attacker held a round seat lid. Replacing it inside the clip with Runway Aleph failed twice and melted the fingers into lumps. Instead, the agent fixed a single frame first with the fingers gripping the flat lid, Kling 4 animated from that frame, and the agent took the 7 hardest frames and added handheld shake and grain to match. Letting the agent do the heavy lifting: - The edit: each Seedance run made 10 to 20 shots, so the film ended up with 239 clips. The agent cut every take at its shot changes using the prompt times, lined each take up with the previz on 19 tracks, and handed me one XML file. I opened it in Resolve and edited purely by deleting what I didn't want, without moving a single clip. - The grade: I have a folder of 645 LUTs. The agent pulled 12 that fit a fight and put them on the same frames side by side, which helped a lot because I only had to pick one and have the agent lift the dark shadows. These saved me a huge amount of time on repetitive work. Also, - Even with all of this, the film still has a bunch of continuity bugs: a door that gets kicked open and is closed again a few shots later, a towel back on its rail, a lid that changes shape. - For me, continuity is still the hardest part of making an AI film.
Wanted to try super fast cutting on an AI fight scene in a extremely simple setting.. 54 secs, 57 shots, the shortest one is 5 frames Most difficult part apparently is planning and editing, not generation... Blocked the whole thing in Blender first with box people, then Seedance. The only software I touched is Davinci Resolve, everything else was done inside @startvoyager, such an efficient way. My theory is, we should only talk to agents like peers, we shouldn't have to learn how to prompt, how to connect nodes, only thing that matters should be creativity itself
19
10
3
68
4,930
Sam E retweeted
Took vork for a spin. Claude Code built this in one session: 1,000,000 zombies, each with its own GPU-simulated steering, crowd avoidance and ragdoll flight. Bullet time from a sim clock written into the shader. 12 s at 1080p, rendered in 21 min on one RTX 5090.
Dear Claude: you are being held back by human-centric graphics APIs. I've built a 3D and VFX engine where agents can write their own shaders, materials, cameras, physics. The core API fits in ~2,500 tokens. Vork is: • 2x more token-efficient vs Blender, Unreal Engine, Houdini • Up to 24x faster rendering times • Rust and Slang based engine, runs on GPU via Metal/Vulkan Below is a simulation of 1 million birds - each with unique trajectories - rendered in 97 sec on one NVIDIA L4 GPU (Blender EEVEE: 31 minutes)
20
16
2
318
19,144
Omg this vork thing is sick and it actually works, seemed a lil sus at first (im the perfect level of paranoid) took their bird murmuration demo and got opus to remix it underwater, 400k fish + a rigged manta ray, took only 9 mins on my macbook m3 128gb prompt in the reply if you wanna try it
Dear Claude: you are being held back by human-centric graphics APIs. I've built a 3D and VFX engine where agents can write their own shaders, materials, cameras, physics. The core API fits in ~2,500 tokens. Vork is: • 2x more token-efficient vs Blender, Unreal Engine, Houdini • Up to 24x faster rendering times • Rust and Slang based engine, runs on GPU via Metal/Vulkan Below is a simulation of 1 million birds - each with unique trajectories - rendered in 97 sec on one NVIDIA L4 GPU (Blender EEVEE: 31 minutes)
5
3
1
57
5,372
Sam E retweeted
Yes! @instagram Edits desktop + @startvoyager is magic 🤯 Just gave it some clips and asked it to make a 30 sec edit with music, sound effects and text overlay Didn't touch the timeline... Got a fun montage of the halloween party I went to + my friend's DJ set.
a pattern I’m liking: creator apps like Instagram Edits and Roblox Studio shipping with built-in MCPs, so agents can do some of the work while power users can collaborate and continue the work in tools that they're used to
1
3
39
3,002
Dear Claude: you are being held back by human-centric graphics APIs. I've built a 3D and VFX engine where agents can write their own shaders, materials, cameras, physics. The core API fits in ~2,500 tokens. Vork is: • 2x more token-efficient vs Blender, Unreal Engine, Houdini • Up to 24x faster rendering times • Rust and Slang based engine, runs on GPU via Metal/Vulkan Below is a simulation of 1 million birds - each with unique trajectories - rendered in 97 sec on one NVIDIA L4 GPU (Blender EEVEE: 31 minutes)
226
444
155
6,805
695,500