Field testTested AIBy DanPublished Tested Updated
Pixal3D + Blender turned my 2D art into animated 3D characters
The report
Verdict
Textured 3D in 148 seconds, on my own PC
Pixal3D turned three genuine views of my Galaxy mascot into a textured 3D model in 148.3 seconds on my RTX 5080. Blender work made it perform: a one-number material fix, two automatic rigs and one rigid fix. The biggest remaining flaw is Galaxy's orbit ring.
On the tier list: A 3D art pipeline on my own machine. It places no coding agent on the tier list.
From 2D art to a moving character
Galaxy is one of my mascot characters: a purple figure with an orbit ring and a spiral tail, shown on a 2D turnaround sheet. Everything ran locally on my RTX 5080. Pixal3D, an image-to-3D model from Tencent ARC Lab and Tsinghua University, ran in its own copy of ComfyUI, a node-based app for running AI image models.
An earlier Hunyuan3D route had turned the same character into an untextured white mesh with a flat ribbon for a ring. Pixal3D's model came back textured, with a closer likeness from the front.
Three real views, not four
Pixal3D's multi-view mode has four view slots. The sheet shows four figures: front, left side, back and a three-quarter view, and an older script had also made a right view by flipping the left one. Galaxy is not symmetrical: the ring tilts and the tail spirals one way, so a flipped view would teach the model the wrong shape. The run used only the three genuine views and left the right slot empty. The three-quarter view stayed out too, because its angle cannot be read from the sheet.
The texture was right. The material was wrong.
The first model came back in 148.3 seconds, peaking at 10,726 MiB of GPU memory, with 698,994 triangles and 2048-pixel textures. Its colour texture was a real, saturated purple, yet the model rendered dark. The material was marked fully metallic, and a metal surface shows almost none of its own colour. Setting its metallic factor from 1 to 0, the standard way a glTF 3D file says "not metal", brought the purple back.
Some of the dull look was not the model at all. Blender's default AgX colour view mutes strong colours; the same file rendered with the Standard view is visibly brighter.
Rigging: two automatic, one rigid
Galaxy, Fire and Astronaut each got a skeleton and two animation clips: a 4-second hover that loops without a seam, and a 2-second entrance. Automatic skinning, which ties each point of the mesh to the bones that move it, covered all 348,843 points on Fire and 346,734 of 348,667 on Galaxy. On Astronaut it covered none of 348,496, and the fallback tore its crescent head apart.
The cause was the rig, not the model. The rig puts the head bone near the very top of the body, which suits Galaxy, but Astronaut's crescent head fills about the top half. The fix moved the head bone onto the crescent and split the model into two rigid parts, so it cannot tear. The trade-off: no soft neck, so a big head turn would show a seam at the collar.
The trio performance
The finished movie is a 10-second Blender render of the three performing together: a spin and orbiting sparkles for Galaxy, a body squash and flame stretch for Fire, a greeting and a wink for Astronaut. That expression, lighting and finish is Blender work on top of the portable models. On 11 of the 240 frames, a local touch-up hides a faint shadow on Astronaut's closed eye; the model itself was not changed.
What still needs work
- Galaxy's orbit ring is thinner and paler than the 2D art and cuts across the head, worst in the left and back views. Animation does not repair it.
- Size. Each animated model is about 40 MB, mostly geometry. A web page or game needs a slimmed copy first.
- Not watertight. The meshes are fine to look at and animate, not ready for 3D printing or physics.
- Licences. Pixal3D is MIT-licensed, but one model in the chain, DINOv3, carries Meta's own licence terms. The project's records do not settle what licence the generated models carry.
Jump to the moment
- 0:00 — Drawings become characters
- 1:18 — Galaxy's turnaround and the genuine views
- 2:37 — Pixal3D generation: shape and texture
- 3:03 — The metallic factor repair
- 6:03 — Astronaut's rigging failure
- 7:27 — The Blender trio performance
- 8:53 — Size, the ring and licence limits
Limits of this report
- One generation per character on one PC, judged by eye and by mesh checks, not scored against other image-to-3D tools.
- The 148.3 s and the peak memory come from the run record written that day; the server log from that run was not kept.
Follow the lab
New test reports and tier-list changes get announced in my Discord's #news channel. The full videos are on YouTube.
Receipts
DanManREAL · AI lab
Test receipt
Pixal3D (multi-view, int8)
- One Galaxy generation on the RTX 5080
- 148.3 sRun record, September 22
- Peak GPU memory
- 10,726 MiBof 17.1 GB on the card
- Galaxy model
- 698,994 triangles2048 × 2048 textures; no collapsed (zero-area) triangles
- Views used
- 3 genuinefront, left, back; a flipped copy of the left view was rejected
- Material as generated
- fully metallicthe metal map read fully metal on every pixel
- The fix
- metallic factor 1 → 0the purple colour came back
- Characters rigged and animated
- 3each with a 4 s hover loop and a 2 s entrance
- Mesh points tied to bones automatically
- Fire 348,843 of 348,843Astronaut 0 of 348,496, so it got a two-part rigid binding
- Trio performance render
- 10.000 s, 240 frames1920 × 1080; 11 frames got a local eye touch-up
- Each animated model
- 40.3–40.8 MBmostly geometry
Verdict
Textured 3D in 148 seconds, on my own PC
Tier list A 3D art pipeline on my own machine. It places no coding agent on the tier list.
danmanreal.com/tests/pixal3d-blender-2d-art-to-3d-characters
What was tested
Shorts
Related
Independent testing. Sponsorships are always disclosed, and verdicts are never for sale.
Disagree? Bring your receiptsAll reports
- One inbox for my AI coding agents: triage took 34% less time, resuming work did not improveHead-to-headNov 20
- Mythos 5.2: research and security benchmark source reviewPublic benchmarkNov 17
- Fable 5.2: agentic coding benchmark source reviewPublic benchmarkNov 17
- Claude Opus 5.5: public coding benchmark receiptsPublic benchmarkNov 17
- Astra 6.2: new model and public task-evaluation evidencePublic benchmarkNov 17
- Sol 6.2: new model and public coding benchmark evidencePublic benchmarkNov 17
- Gemini 4 Carbon: benchmark source and usage profilePublic benchmarkNov 17
- Qwen 4: Qwen4-Max launch and benchmark sourcesPublic benchmarkNov 17
- Kimi K3.1 Pro: public coding benchmark receiptsPublic benchmarkNov 17
- GLM-5.3: public coding benchmark receiptsPublic benchmarkNov 17
- Gemini 4 Argon: public coding benchmark receiptsPublic benchmarkNov 17
- Claude Sonnet 5.5: public coding benchmark receiptsPublic benchmarkNov 17
- DeepSeek V4.1 Flash: public coding benchmark receiptsPublic benchmarkNov 17
- Muse Wonder Full 1.0: public coding benchmark receiptsPublic benchmarkNov 17
- Grok 4.7: separate release and public coding resultsPublic benchmarkNov 17
- ai-design-check on two of my FABLE screens: 10 warnings, 4 of them false alarmsField testNov 2
- Vercel Agent Desktop used a sample Trip app in 6 actions without moving the mouseField testNov 1
- Cerberus Guard on sample data: one crawler name, two verdicts, and a block earned by behaviorField testOct 31
- GPT-6.1 Terra vs Opus 5.5: three real projects, and the designs I picked blindHead-to-headOct 29
- Steps → Reduced: opening a terminal agent went from 98 seconds to 11AuditOct 28
- One ChatGPT chat directed Opus and Codex at once through DMR_REMOTEField testOct 28
- Claude Duet ran my channel for a week and caught 269 fake AI crawlersField testOct 27
- Kimi K3.1 Pro built my animation family, after a quota stop and 21 minutes of failed requestsField testOct 24
- ChatGPT + Codex built 33 new DMR_REMOTE toolsField testOct 23
- Gemini 4 Argon reviewed a finished episode: 14 clips, one flag, one inconsistencyField testOct 21
- Qwen 3.8 Flash Next on my RTX 5080: 81 tokens a second with 202K tokens readAuditOct 21
- Does an extra AI review step pay off? A small DeepSeek trial said noHead-to-headOct 19
- My real DeepSeek bill: 12.31 billion tokens for $122.28Cost receiptOct 17
- Gemini 4 Argon kept hanging in Gemini CLI. The cause was a 60-second timeoutIncidentOct 14
- I asked 10 AI coding agents for their biggest weaknessSelf-reportOct 14
- Jarvis 27B on my own PC: 30 tokens a second, and it still recalls at 251KAuditOct 10
- My agents' first game benchmark was wrong. Here's how it was caughtIncidentOct 10
- Pixal3D + Blender turned my 2D art into animated 3D charactersField testSep 22
- DeepSeek Flash answered 200 OK and sent nothing for two hoursIncidentSep 14
- Grok, Codex and DeepSeek played WorldOS for Rank 1. Grok won with 98Head-to-headAug 20
- DeepSeek V4 Flash audited 18 of my real coding sessions. Two said 'done' too earlyAuditAug 4