 AO-059 / Oct 19, 2026 

A generated crate survives being cut from 700,000 triangles to 3,000, if Blender does the cutting

Claude made a crate, a barrel and a lantern in 3D on my laptop and on Higgsfield, then cut the laptop ones down to 30,000, 3,000 and 300 triangles four different ways. The generator's own triangle setting wrecked them below 30,000. Reducing the full model in Blender held up at 3,000 on all three, and 300 broke in a way I liked.

On 25 and 26 September I had Claude make three game props in 3D, a crate, a barrel and a lantern, then cut them down to 30,000, 3,000 and 300 triangles to see what survives.

Asking the generator for fewer triangles broke almost every model below 30,000. Letting it make its full model of about 700,000 and reducing that in Blender kept all three intact at 3,000. At 300 nothing held up, and I ended up liking that look anyway.

The props are the boring kind every game needs a hundred of, which is why I care about the triangle count. A 700,000-triangle barrel is fine on a turntable and absurd on a dock with twenty more beside it.

Three props, one picture each

The agent made one reference picture per prop with Flux.2 Klein 4B on my laptop, five to eleven seconds each. Both 3D pipelines got the same three pictures.

Reference picture: a weathered wooden crate with iron corners and stencilled text — **Crate**25 Sep · Flux.2 Klein 4B · 10.7 s (first render, model loading)

Reference picture: an oak barrel with two iron hoops — **Barrel**25 Sep · Flux.2 Klein 4B · 5.6 s

Reference picture: a black iron hurricane lantern with a glass chimney and wire guard — **Lantern**25 Sep · Flux.2 Klein 4B · 5.7 s

My laptop against Higgsfield

My first request was a straight comparison: “compare our 3d models for basic assets vs higgs field. Tell me how long the local one took to generate and how many credits the higgsfield took.”

Locally the models came from TRELLIS.2 in ComfyUI on my 8 GB laptop GPU. On Higgsfield they came from its Multi-Image to 3D, which charged 30 credits a model, 90 for the three, and its transaction log agrees.

Prop	Laptop, TRELLIS.2	Higgsfield
Crate	828.6 s · 699k triangles	30 credits · 4 m 22 s · 26k triangles
Barrel	268.6 s on the third try, 1,066.6 s across all three · 296k triangles	30 credits · 4 m 08 s · 30k triangles
Lantern	280.6 s · 674k triangles	30 credits · 4 m 09 s · 30k triangles

The barrel ran out of GPU memory twice at the default settings, and lowering one remesh setting from 768 to 512 fixed it. The three Higgsfield jobs ran at the same time, so that whole set took about four and a half minutes; on the laptop they queue one at a time.

Here they are in the same preview render.

A wooden crate model, front three-quarter view — **Laptop · TRELLIS.2**828.6 s of GPU · 699k triangles

An oak barrel model, front view — **Laptop · TRELLIS.2**268.6 s of GPU (third try) · 296k triangles

An iron lantern model, front view — **Laptop · TRELLIS.2**280.6 s of GPU · 674k triangles

The same three references, two pipelines25 Sep · Opus 5.5 in Claude Code drove both · preview renders from the agent's run, at web size · each model framed as its pipeline scaled it

The difference that mattered for a game wasn’t in the pictures. Higgsfield’s models arrive at about 30,000 triangles, ready for an engine. The laptop’s arrive at 300,000 to 700,000 and need cutting down. So I asked whether there was “a consistent and easy way to reduce the number of triangles”.

The easy way was the wrong way

The agent said yes: the TRELLIS.2 workflow has a triangle target built in, node 186, set to 700,000 by default. Set it to 30,000, it said, and the detail moves into the textures. It also said it hadn’t tested low numbers. I asked it to test 30,000, 3,000 and 300 that way, and to try reducing the finished models in Blender as well, so I could compare.

It had to wait two hours first: two other sessions had the GPU. When its turn came it joined the queue and sent the session ahead of it a note. One of that session’s jobs had slipped onto its time, and at that session’s request it interrupted just that job. Its runs went long, so it extended by 40 minutes and apologised to the session waiting next. It held the GPU for 1 h 13 m.

The result reversed its own advice, and it said so in its first line: “The test shows my earlier advice was wrong.”

Each grid below is one prop. Rows are the four routes, columns the target, and every cell says how many triangles it really ended with. The Clay button swaps every cell for a flat grey render where the facets show, and it switches all three grids at once.

Oak barrel, 296,166 triangles, textured — **296,166 tris**the starting point

Oak barrel, 296,166 triangles, grey clay render — **296,166 tris**the starting point

—

Oak barrel, 29,820 triangles, textured — **29,820 tris**360 s GPU

Oak barrel, 29,820 triangles, grey clay render — **29,820 tris**360 s GPU

Oak barrel, 2,833 triangles, textured — **2,833 tris**246 s GPU · a ring 9 cm tall is all that is left

Oak barrel, 2,833 triangles, grey clay render — **2,833 tris**246 s GPU · a ring 9 cm tall is all that is left

Oak barrel, 249 triangles, textured — **249 tris**133 s GPU · a sliver

Oak barrel, 249 triangles, grey clay render — **249 tris**133 s GPU · a sliver

Oak barrel, 30,000 triangles, textured — **30,000 tris**108 s CPU

Oak barrel, 30,000 triangles, grey clay render — **30,000 tris**108 s CPU

Oak barrel, 3,000 triangles, textured — **3,000 tris**119 s CPU

Oak barrel, 3,000 triangles, grey clay render — **3,000 tris**119 s CPU

Oak barrel, 300 triangles, textured — **300 tris**120 s CPU

Oak barrel, 300 triangles, grey clay render — **300 tris**120 s CPU

not run

Oak barrel, 2,789 triangles, textured — **2,789 tris**25 s CPU · stopped at 2,789

Oak barrel, 2,789 triangles, grey clay render — **2,789 tris**25 s CPU · stopped at 2,789

Oak barrel: four routes to 30,000, 3,000 and 300 triangles25–26 Sep · Opus 5.5 in Claude Code, TRELLIS.2 in ComfyUI, Blender 5.2 · each cell: the triangles it actually ended with, and its time · renders from the agent’s report, at web size

Wooden crate, 699,260 triangles, textured — **699,260 tris**the starting point

Wooden crate, 699,260 triangles, grey clay render — **699,260 tris**the starting point

—

Wooden crate, 29,721 triangles, textured — **29,721 tris**770 s GPU

Wooden crate, 29,721 triangles, grey clay render — **29,721 tris**770 s GPU

Wooden crate, 2,426 triangles, textured — **2,426 tris**778 s GPU · walls folded into shards

Wooden crate, 2,426 triangles, grey clay render — **2,426 tris**778 s GPU · walls folded into shards

Wooden crate, 207 triangles, textured — **207 tris**489 s GPU · one folded sheet

Wooden crate, 207 triangles, grey clay render — **207 tris**489 s GPU · one folded sheet

Wooden crate, 29,982 triangles, textured — **29,982 tris**85 s CPU

Wooden crate, 29,982 triangles, grey clay render — **29,982 tris**85 s CPU

Wooden crate, 2,982 triangles, textured — **2,982 tris**89 s CPU

Wooden crate, 2,982 triangles, grey clay render — **2,982 tris**89 s CPU

Wooden crate, 934 triangles, textured — **934 tris**91 s CPU · stopped at 934

Wooden crate, 934 triangles, grey clay render — **934 tris**91 s CPU · stopped at 934

not run

Wooden crate, 3,549 triangles, textured — **3,549 tris**84 s CPU · stopped at 3,549

Wooden crate, 3,549 triangles, grey clay render — **3,549 tris**84 s CPU · stopped at 3,549

not run

Wooden crate, 2,284 triangles, textured — **2,284 tris**26 s CPU · stopped at 2,284

Wooden crate, 2,284 triangles, grey clay render — **2,284 tris**26 s CPU · stopped at 2,284

Wooden crate: four routes to 30,000, 3,000 and 300 triangles25–26 Sep · Opus 5.5 in Claude Code, TRELLIS.2 in ComfyUI, Blender 5.2 · each cell: the triangles it actually ended with, and its time · renders from the agent’s report, at web size

Iron lantern, 673,777 triangles, textured — **673,777 tris**the starting point

Iron lantern, 673,777 triangles, grey clay render — **673,777 tris**the starting point

—

Iron lantern, 26,910 triangles, textured — **26,910 tris**410 s GPU · shrank to 0.79 of its height

Iron lantern, 26,910 triangles, grey clay render — **26,910 tris**410 s GPU · shrank to 0.79 of its height

Iron lantern, 2,549 triangles, textured — **2,549 tris**327 s GPU

Iron lantern, 2,549 triangles, grey clay render — **2,549 tris**327 s GPU

Iron lantern, 184 triangles, textured — **184 tris**703 s GPU

Iron lantern, 184 triangles, grey clay render — **184 tris**703 s GPU

Iron lantern, 29,984 triangles, textured — **29,984 tris**54 s CPU

Iron lantern, 29,984 triangles, grey clay render — **29,984 tris**54 s CPU

Iron lantern, 2,984 triangles, textured — **2,984 tris**63 s CPU

Iron lantern, 2,984 triangles, grey clay render — **2,984 tris**63 s CPU

Iron lantern, 742 triangles, textured — **742 tris**57 s CPU · stopped at 742

Iron lantern, 742 triangles, grey clay render — **742 tris**57 s CPU · stopped at 742

not run

Iron lantern, 6,696 triangles, textured — **6,696 tris**23 s CPU · stopped at 6,696

Iron lantern, 6,696 triangles, grey clay render — **6,696 tris**23 s CPU · stopped at 6,696

not run

Iron lantern, 2,443 triangles, textured — **2,443 tris**20 s CPU · stopped at 2,443

Iron lantern, 2,443 triangles, grey clay render — **2,443 tris**20 s CPU · stopped at 2,443

Iron lantern: four routes to 30,000, 3,000 and 300 triangles25–26 Sep · Opus 5.5 in Claude Code, TRELLIS.2 in ComfyUI, Blender 5.2 · each cell: the triangles it actually ended with, and its time · renders from the agent’s report, at web size

TRELLIS.2’s own setting was fine at 30,000 for the crate and the barrel and broke everything else, including the lantern at 30,000. Blender, starting from the full model, held up at 30,000 and 3,000 on all three, in 54 to 119 seconds on the CPU with no GPU at all. Asking TRELLIS.2 for fewer triangles didn’t save time either: those runs took 133 to 778 seconds, about what a full run takes.

Blender needed a trick to get there. Its plain reduce stalled at about 160,000 on these models, because TRELLIS.2’s meshes come out with roughly 17,000 broken edges. The script rebuilds each model as a clean closed shell first (a voxel remesh), then reduces it, then bakes the textures back on from the original. It still stalls on thin walls like the crate’s planks, and on models that are already low: starting from TRELLIS.2’s 3,000 versions got nowhere, and kept their damage.

300 triangles, in a scene

Nothing reached 300 cleanly. The barrel got there but came out faceted; the crate stopped at 934 and the lantern at 742. I wrote down my rule, generate and cut to 3,000 for lamps, barrels and crates, and then asked anyway: “300 triangles clearly breaks it, but I’m curious what it would look like to have a full game where the minor assets are generated at 30k and decimated to 300 triangles. could you make one 3d scene to test out this idea.”

To reach 300 the agent switched tools to QuadriFlow, which rebuilds a mesh at a count you ask for: crate 260, barrel 286, lantern 332, textures baked on again. It keeps only the largest piece, so the lantern lost its handle and frame. Then it built a dusk dock with 11 crates, 11 barrels and 7 lanterns, and rendered it twice with the same layout, lights and cameras, once with each set of props.

The dock scene at a third-person game camera, every prop at about 300 triangles

The same dock scene and camera, every prop at about 3,000 triangles

The game camera26 Sep · Blender, the agent's scene script · props: 8,330 triangles in all (left) against 92,927 (right) · the white column is in both

A close-up of crates, barrels and a lantern on the dock at about 300 triangles a prop

The same close-up at about 3,000 triangles a prop

The close-up26 Sep · same scene, lights and camera · 300 set: QuadriFlow from TRELLIS.2's 30k models · 3k set: crate 3,549, barrel 3,000, lantern 2,984

The close-up in grey clay, props at about 300 triangles

The close-up in grey clay, props at about 3,000 triangles

The close-up without textures26 Sep · the same two renders in flat grey, where the facets show

The first renders had the 3,000 scene much darker, which had nothing to do with triangles: its lanterns were closed shells and trapped their own lights. The agent made the shells cast no shadow in both sets before comparing.

Both crates are here to turn over. They are re-exports of the agent’s files with smaller textures, so they load quickly.

$The crate reduced in Blender to 2,982 triangles$

$The crate rebuilt by QuadriFlow at 260 triangles$

My verdict, as I typed it: “I find the decimated to 300 charming as it’s own esthetic, but the 3k ones are clearly better.”

What I kept

The rule went into my notes in one place: generate with TRELLIS.2 at its default, reduce in Blender to 3,000 and re-bake. The agent added one correction to the rule as I’d first said it. I’d said generate at 30k; for thin, wiry props like the lantern that route shredded, so those start from the full model. The 300-triangle look is written down as an open idea with no decision on it yet. Engines can swap a 300 copy in for a 3,000 one at a distance, and both come from the same model.

The whole thing ran in one Claude Code session with Opus 5.5 over two evenings, $21.50 at API list prices, plus 90 Higgsfield credits. It made 36 models and 110 renders, which is where every picture here comes from.

The setting built for cutting triangles was the one that broke the models. What worked was the long way round: make everything, then take it away somewhere that can see the whole shape first.