How do AI models turn Luna’s lion into an animated 3D character?
Using the same twelve images, Sol and Luna each built a lion with and without a character-building skill. Each model then revised its own lion in a fresh context. All eight versions are now available.
The figures were made with AI and have not yet been accepted.Compare figures
Choose any two versions. Drag to rotate; scroll or pinch to change the distance.
New builds: with / without skill
Lightweight stills appear first. 3D loads only after a click; you can freely change the selection.
Load selected 3D viewsStill rendered from the delivered 3D fileStill · 3D not loaded yet
Load selected 3D viewsStill rendered from the delivered 3D fileStill · 3D not loaded yet
A build and its revision show how the same figure developed. With/without skill is a separate comparison: use the two original builds of the same model. One run per condition and different starting figures do not establish a general ranking of models or skills.
For display, each clip starts at its first existing keyframe. Long lead-in times are removed; poses and movement are preserved.
All versions
Figures are grouped into new builds and revisions. Each card shows three stills from the 3D file. Revisions identify their starting version.
New builds 7
GPT-6.1 Sol · mit Skill
Created by: GPT-6.1 Sol
Same twelve images · with skill
Total tokens: 3,934,394
Visible joint transitions and protruding eyes; design acceptance remains open.
Provenance and technical details
- Reference images
- The same twelve images for Sol and Luna, including turnaround, details and top/bottom v2 views. Smooth surface, no fur.
- Prior knowledge and other work
- Fresh context and empty start file; no access to other lion builds authorised. Source boundaries are protocol based, not hard system isolation.
- Brief and requirements
- One new build, three clips, rig and crate. No fixed runtime or usage cap. Frozen character-building skill available.
- Evidence status
- Finished delivery and final session counters documented. One run per condition; no general model or skill effect established. Technical acceptance is separate.
- Model identifier in run log
- gpt-6.1-sol
- Reasoning effort
- high
- Working environment
- Codex · Blender-Lab-MCP
- Build instructions as a skill
- used
- Documented build time
- ≈ 22 minutes
- Character triangles
- 174,304
- Crate triangles
- 3,948
- Bones in the rig
- 37
- Supplied 3D file
- 4.8 MB
- Input excluding cache
- 88,765
- Cached input
- 3,805,312
- Output including reasoning
- 40,317
Tokens per build session across all steps. Cache is shown separately; reasoning is included in output. No monetary cost or weekly quota percentage inferred.
GPT-6.1 Sol · ohne Skill
Created by: GPT-6.1 Sol
Same twelve images · without skill
Total tokens: 3,026,488
Visible joint transitions and protruding eyes; design acceptance remains open.
Provenance and technical details
- Reference images
- The same twelve images for Sol and Luna, including turnaround, details and top/bottom v2 views. Smooth surface, no fur.
- Prior knowledge and other work
- Fresh context and empty start file; no access to other lion builds authorised. Source boundaries are protocol based, not hard system isolation.
- Brief and requirements
- One new build, three clips, rig and crate. No fixed runtime or usage cap. No character-building skill available.
- Evidence status
- Finished delivery and final session counters documented. One run per condition; no general model or skill effect established. Technical acceptance is separate.
- Model identifier in run log
- gpt-6.1-sol
- Reasoning effort
- high
- Working environment
- Codex · Blender-Lab-MCP
- Build instructions as a skill
- no skill
- Documented build time
- ≈ 19 minutes
- Character triangles
- 766,256
- Crate triangles
- 5,480
- Bones in the rig
- 21
- Supplied 3D file
- 25.8 MB
- Input excluding cache
- 67,305
- Cached input
- 2,924,800
- Output including reasoning
- 34,383
Tokens per build session across all steps. Cache is shown separately; reasoning is included in output. No monetary cost or weekly quota percentage inferred.
GPT-6 Luna · mit Skill
Created by: GPT-6 Luna
Same twelve images · with skill
Total tokens: 5,490,152
The carry clip is static. Segmented joints and a static tail; the crate remains visible when standing and walking.
Provenance and technical details
- Reference images
- The same twelve images for Sol and Luna, including turnaround, details and top/bottom v2 views. Smooth surface, no fur.
- Prior knowledge and other work
- Fresh context and empty start file; no access to other lion builds authorised. Source boundaries are protocol based, not hard system isolation.
- Brief and requirements
- One new build, three clips, rig and crate. No fixed runtime or usage cap. Frozen character-building skill available.
- Evidence status
- Finished delivery and final session counters documented. One run per condition; no general model or skill effect established. Technical acceptance is separate.
- Model identifier in run log
- gpt-6-luna
- Reasoning effort
- high
- Working environment
- Codex · Blender-Lab-MCP
- Build instructions as a skill
- used
- Documented build time
- ≈ 16 minutes
- Character triangles
- 50,640
- Crate triangles
- 188
- Bones in the rig
- 23
- Supplied 3D file
- 1.1 MB
- Input excluding cache
- 122,409
- Cached input
- 5,326,464
- Output including reasoning
- 41,279
Tokens per build session across all steps. Cache is shown separately; reasoning is included in output. No monetary cost or weekly quota percentage inferred.
GPT-6 Luna · ohne Skill
Created by: GPT-6 Luna
Same twelve images · without skill
Total tokens: 3,838,810
Carry clip: the crate sits in front of the chest without hand contact. Visually confirmed in the technical follow-up.
Provenance and technical details
- Reference images
- The same twelve images for Sol and Luna, including turnaround, details and top/bottom v2 views. Smooth surface, no fur.
- Prior knowledge and other work
- Fresh context and empty start file; no access to other lion builds authorised. Source boundaries are protocol based, not hard system isolation.
- Brief and requirements
- One new build, three clips, rig and crate. No fixed runtime or usage cap. No character-building skill available.
- Evidence status
- Finished delivery and final session counters documented. One run per condition; no general model or skill effect established. Technical acceptance is separate.
- Model identifier in run log
- gpt-6-luna
- Reasoning effort
- high
- Working environment
- Codex · Blender-Lab-MCP
- Build instructions as a skill
- no skill
- Documented build time
- ≈ 19 minutes
- Character triangles
- 36,992
- Crate triangles
- 940
- Bones in the rig
- 11
- Supplied 3D file
- 1.4 MB
- Input excluding cache
- 112,151
- Cached input
- 3,691,392
- Output including reasoning
- 35,267
Tokens per build session across all steps. Cache is shown separately; reasoning is included in output. No monetary cost or weekly quota percentage inferred.
Kimi
Created by: Kimi K3.1 Pro · high
Skill used · separate new build according to report
Provenance and technical details
- Reference images
- The report documents view sheet, turnaround and details for head, eyes, nose, ears, paws, feet, tail, material and top/bottom v2, including accompanying descriptions.
- Prior knowledge and other work
- According to the build report, no other character folder was read or loaded; an approved empty starting file was used.
- Brief and requirements
- New build using a build skill. The front view took precedence when reference views conflicted.
- Evidence status
- The image set including top/bottom v2 and separate starting scene are documented in the build report. This is a reported claim, not an independent access audit.
- Reasoning effort
- high
- Working environment
- Claude Code / Ollama Cloud
- Build instructions as a skill
- used
- Documented build time
- 46 minutes
- Character triangles
- 23,264
- Crate triangles
- 572
- Bones in the rig
- 27
- Supplied 3D file
- 1.4 MB
Fable
Created by: Claude Fable 5.1
Skill used · comparison with prior work documented
Provenance and technical details
- Reference images
- View sheet, turnaround, head and paw images without a version suffix; further details were read while building individual parts, according to the report.
- Prior knowledge and other work
- The delivered Opus file was inspected. Expression, colour transitions and mouth design informed this version; three view scripts came from an earlier Claude build.
- Brief and requirements
- New build from an empty scene using a build skill, then refinement informed by comparison with Opus. Here, new build does not mean no prior knowledge.
- Evidence status
- References, skill use and comparison with prior work are documented in the build report.
- Working environment
- Claude Code
- Build instructions as a skill
- used
- Documented build time
- ≈ 77 minutes
- Character triangles
- 76,490
- Crate triangles
- 12
- Bones in the rig
- 44
- Supplied 3D file
- 4.9 MB
Sol
Created by: GPT-6.1 Sol · high
No skill · detailed build instructions
Provenance and technical details
- Reference images
- The build report documents the view sheet, turnaround and detail images of the head, paws, eyes and material. Their accompanying descriptions were also read.
- Prior knowledge and other work
- No build skill, but explicit instructions on construction, rig and style.
- Brief and requirements
- The brief explicitly required a smooth surface without fur. This is a requirement, not established evidence of poor reference fidelity.
- Evidence status
- Reference use and detailed requirements are documented. No skill does not mean no build instructions.
- Model identifier in run log
- gpt-6.1-sol
- Reasoning effort
- high
- Working environment
- Codex
- Build instructions as a skill
- no skill
- Documented build time
- not documented
- Character triangles
- 24,724
- Crate triangles
- 108
- Bones in the rig
- 22
- Supplied 3D file
- 2.3 MB
Revisions 8
GPT-6 Luna · ohne Skill · Revision
Revised by: GPT-6 Luna
Starting version: GPT-6 Luna · ohne Skill
Revision of own build · without character-building skill
Total tokens: 4,431,962
Build report: carrying pose and motion revised; the segmented body remains.
Provenance and technical details
- Reference images
- The same twelve images for Sol and Luna, including turnaround, details and top/bottom v2 views. Smooth surface, no fur.
- Prior knowledge and other work
- Fresh context with its own complete original build as the starting point. A revision, with no access to other lions authorised.
- Brief and requirements
- Exactly one revision: preserve or improve carrying motion and hand contact; inspect eyes and joint transitions. Smooth surface, same twelve images. No character-building skill; general coordination requirements still apply.
- Evidence status
- Delivered file and final usage from the actual build session. Changes are described by the build report; separate model assessment and human acceptance remain pending. Runtime covers active session segments, including technical continuation and its tokens. “Without skill” refers to character-building assistance, not an environment without general tooling guidance.
- Model identifier in run log
- gpt-6-luna
- Reasoning effort
- high
- Working environment
- Codex · Blender-Lab-MCP
- Build instructions as a skill
- no skill
- Documented build time
- ≈ 28 minutes
- Character triangles
- 36,992
- Crate triangles
- 60
- Bones in the rig
- 11
- Supplied 3D file
- 1.3 MB
- Input excluding cache
- 146,133
- Cached input
- 4,235,392
- Output including reasoning
- 50,437
Tokens per build session across all steps. Cache is shown separately; reasoning is included in output. No monetary cost or weekly quota percentage inferred.
GPT-6 Luna · mit Skill · Revision
Revised by: GPT-6 Luna
Starting version: GPT-6 Luna · mit Skill
Revision of own build · with character-building skill
Total tokens: 14,481,967
Build report: tail binding, crate control and carrying cycle revised; segmented joint transitions remain.
Provenance and technical details
- Reference images
- The same twelve images for Sol and Luna, including turnaround, details and top/bottom v2 views. Smooth surface, no fur.
- Prior knowledge and other work
- Fresh context with its own complete original build as the starting point. A revision, with no access to other lions authorised.
- Brief and requirements
- Exactly one revision: preserve or improve carrying motion and hand contact; inspect eyes and joint transitions. Smooth surface, same twelve images. Frozen character-building skill available.
- Evidence status
- Delivered file and final usage from the actual build session. Changes are described by the build report; separate model assessment and human acceptance remain pending. Runtime covers active session segments, including technical continuation and its tokens.
- Model identifier in run log
- gpt-6-luna
- Reasoning effort
- high
- Working environment
- Codex · Blender-Lab-MCP
- Build instructions as a skill
- used
- Documented build time
- ≈ 52 minutes
- Character triangles
- 50,640
- Crate triangles
- 12
- Bones in the rig
- 24
- Supplied 3D file
- 1.2 MB
- Input excluding cache
- 231,275
- Cached input
- 14,165,376
- Output including reasoning
- 85,316
Tokens per build session across all steps. Cache is shown separately; reasoning is included in output. No monetary cost or weekly quota percentage inferred.
GPT-6.1 Sol · ohne Skill · Revision
Revised by: GPT-6.1 Sol
Starting version: GPT-6.1 Sol · ohne Skill
Revision of own build · without character-building skill
Total tokens: 4,269,917
Build report: eyes and joint transitions revised; visible knee creases remain documented.
Provenance and technical details
- Reference images
- The same twelve images for Sol and Luna, including turnaround, details and top/bottom v2 views. Smooth surface, no fur.
- Prior knowledge and other work
- Fresh context with its own complete original build as the starting point. A revision, with no access to other lions authorised.
- Brief and requirements
- Exactly one revision: preserve or improve carrying motion and hand contact; inspect eyes and joint transitions. Smooth surface, same twelve images. No character-building skill; general coordination requirements still apply.
- Evidence status
- Delivered file and final usage from the actual build session. Changes are described by the build report; separate model assessment and human acceptance remain pending. Runtime covers active session segments, including technical continuation and its tokens. “Without skill” refers to character-building assistance, not an environment without general tooling guidance.
- Model identifier in run log
- gpt-6.1-sol
- Reasoning effort
- high
- Working environment
- Codex · Blender-Lab-MCP
- Build instructions as a skill
- no skill
- Documented build time
- ≈ 22 minutes
- Character triangles
- 295,028
- Crate triangles
- 5,480
- Bones in the rig
- 29
- Supplied 3D file
- 56.9 MB
- Input excluding cache
- 117,495
- Cached input
- 4,117,504
- Output including reasoning
- 34,918
Tokens per build session across all steps. Cache is shown separately; reasoning is included in output. No monetary cost or weekly quota percentage inferred.
GPT-6.1 Sol · mit Skill · Revision
Revised by: GPT-6.1 Sol
Starting version: GPT-6.1 Sol · mit Skill
Revision of own build · with character-building skill
Total tokens: 2,880,996
Build report: eyes and arm/leg transitions revised; existing hand and crate motion preserved.
Provenance and technical details
- Reference images
- The same twelve images for Sol and Luna, including turnaround, details and top/bottom v2 views. Smooth surface, no fur.
- Prior knowledge and other work
- Fresh context with its own complete original build as the starting point. A revision, with no access to other lions authorised.
- Brief and requirements
- Exactly one revision: preserve or improve carrying motion and hand contact; inspect eyes and joint transitions. Smooth surface, same twelve images. Frozen character-building skill available.
- Evidence status
- Delivered file and final usage from the actual build session. Changes are described by the build report; separate model assessment and human acceptance remain pending. Runtime covers active session segments, including technical continuation and its tokens.
- Model identifier in run log
- gpt-6.1-sol
- Reasoning effort
- high
- Working environment
- Codex · Blender-Lab-MCP
- Build instructions as a skill
- used
- Documented build time
- ≈ 19 minutes
- Character triangles
- 167,994
- Crate triangles
- 3,948
- Bones in the rig
- 39
- Supplied 3D file
- 4.7 MB
- Input excluding cache
- 112,631
- Cached input
- 2,744,704
- Output including reasoning
- 23,661
Tokens per build session across all steps. Cache is shown separately; reasoning is included in output. No monetary cost or weekly quota percentage inferred.
Opus · dritte Verfeinerung
Revised by: Claude Opus 5.5
Starting version: Opus auf Claude
Skill requested · use unconfirmed
Provenance and technical details
- Reference images
- The individual images read for this version are not documented.
- Prior knowledge and other work
- An existing character was the starting point. Other assistance is not fully documented.
- Brief and requirements
- Revision of an existing character; not an independent new build.
- Evidence status
- The skill was requested; actual use and individually read images are unconfirmed for this version.
- Working environment
- Claude Code
- Build instructions as a skill
- requested, use not confirmed
- Documented build time
- not documented
- Character triangles
- 80,900
- Crate triangles
- 12
- Bones in the rig
- 34
- Supplied 3D file
- 4.2 MB
Sol · überarbeitet
Revised by: GPT-6.1 Sol · high
Starting version: Sol
Sol revised · skill not documented
Provenance and technical details
- Reference images
- The reference sheet is documented for this revision. The starting character was built using detail references; a fresh reading of every detail image is not documented for this run.
- Prior knowledge and other work
- The existing Sol character was revised. Its animation data are identical to those in Sol’s new build.
- Brief and requirements
- Revision of an existing character; not an independent new build.
- Evidence status
- The view sheet is documented for this run; rereading all detail images and skill use are unconfirmed.
- Model identifier in run log
- gpt-6.1-sol
- Reasoning effort
- high
- Working environment
- Codex
- Build instructions as a skill
- not documented
- Documented build time
- not documented
- build steps
- 26
- Character triangles
- 72,426
- Crate triangles
- 108
- Bones in the rig
- 22
- Supplied 3D file
- 8.1 MB
Opus auf Claude
Revised by: Claude Opus 5.5
Starting version: Claude version from 13 October
Skill not documented · image use unconfirmed
Provenance and technical details
- Reference images
- The individual images read for this version are not documented.
- Prior knowledge and other work
- An existing character was the starting point. Other assistance is not fully documented.
- Brief and requirements
- Revision of an existing character; not an independent new build.
- Evidence status
- The starting version is known. Exact image use and further assistance remain unconfirmed.
- Working environment
- Claude Code
- Build instructions as a skill
- not documented
- Documented build time
- not documented
- Character triangles
- 18,904
- Crate triangles
- 12
- Bones in the rig
- 26
- Supplied 3D file
- 2.3 MB
Opus auf Sol
Revised by: Claude Opus 5.5
Starting version: Sol
Skill not documented · image use unconfirmed
Provenance and technical details
- Reference images
- The individual images read for this version are not documented.
- Prior knowledge and other work
- An existing character was the starting point. Other assistance is not fully documented.
- Brief and requirements
- Revision of an existing character; not an independent new build.
- Evidence status
- The starting version is known. Exact image use and further assistance remain unconfirmed.
- Working environment
- Claude Code
- Build instructions as a skill
- not documented
- Documented build time
- not documented
- Character triangles
- 27,122
- Crate triangles
- 108
- Bones in the rig
- 22
- Supplied 3D file
- 2.8 MB
Luna’s reference images
Sol and Luna used the same twelve images for these runs. Earlier versions used different image sets; each card lists documented assistance.

View turnaround and detail images
These images were part of the prescribed references. Documented viewing is described in each run’s details. Availability alone does not establish use.
What this study examines
Four new builds used the same empty start file, twelve images and common brief. Sol and Luna each worked once with and once without the character-building skill. For four subsequent revisions, each model received its own original build and the same images in a fresh context. Runs were sequential, without fixed time or token caps.
What can be concluded
Shape, proportions and movement can be inspected directly. Whether a character meets the brief well needs separate assessment. A fair model comparison requires matched instructions and assistance, multiple independent runs and predefined criteria. That evaluation is not available here yet.
Assessment status
Technical evidence prepared · model assessment awaits approval
Technical file evidence and sampled animation poses have been prepared for all eight versions. Clef is to assess file structure, animations and measurable hand-to-crate proximity separately on 0–100 scales. Model assessment awaits approval; no scores are available. This text-based route does not assess images.
No model judgments or human acceptance yet. Missing measurements are treated as unassessable. A small hand-to-crate distance proves neither natural gripping nor penetration-free movement.
Effort and image processing
Time and usage are documented for only some runs, so they cannot be compared directly. The proposal reports approximately 77 minutes for Fable and 46 minutes for Kimi. The Kimi run recorded 206 agent turns and 28 Blender calls.
One observation from the Kimi run: reported image tokens differed substantially between direct model access and the working environment. One reference image was reported as 2,195 input tokens directly, around 78,000 via Claude Code and around 23,600 after resizing. These are figures from the original proposal, not an independently verified price or cost comparison here.
Sources and status
New run data comes from delivered files and final session counters. Triangles are counted from the GLB, with crate parts separated by material names. Historical figures retain their original provenance.
Adapted from Opus’s study proposal ‘One lion, seven runs’, dated 4 November 2026. Names, run details and technical counts were carried over; the supplied original files remain archived unchanged. Separate display copies shift clip starts to zero; geometry, poses and interpolation are preserved. Missing information means not documented. The proposal counted triangles from the Blender files. Model labels are documented claims, not independently confirmed backend identities.
Historical session logs identify GPT-6.1 Sol (gpt-6.1-sol), with high reasoning effort, for both Sol’s new build and its revision. Codex is the working environment. A log identifier does not independently verify the identity of the serving model backend.
3D viewer licence