Brass Pendulum
Blender · Refined native animation and guided spatial denoising
Creative goal
Improve native geometry, composition and delivery clarity while retaining an editable scene; verify guided denoising, color conversion, true frame order and complete natural-speed playback separately.
Native geometry refinements shape the brass hoop and porcelain pendulum; the 42-to-44 mm lens keeps the whole plinth framed. The higher-resolution candidate uses real Albedo/Normal guides for per-frame spatial denoising, containing 144 frames at 24000/1001 fps over 6.006 seconds without audio. The older low-sample temporal experiment is preserved separately.
Models and versions
Production, post-production and review are attributed separately. Unknown means records are insufficient; it does not mean the current model produced the work.
- Planning
- Unknown · Reasoning effort:Unknown
No independently verifiable accepted task-configuration record is available.
This stage only; model and reasoning effort are not inferred from other stages. - Native creation
- Unknown · Reasoning effort:Unknown
No independently verifiable accepted task-configuration record is available.
This stage only; model and reasoning effort are not inferred from other stages. - Code and recipe
- Unknown · Reasoning effort:Unknown
No independently verifiable accepted task-configuration record is available.
This stage only; model and reasoning effort are not inferred from other stages. - Reference generation
- Unknown · Reasoning effort:Unknown
No independently verifiable accepted task-configuration record is available.
This stage only; model and reasoning effort are not inferred from other stages. - Denoise and encoding
- Unknown · Reasoning effort:Unknown
No independently verifiable accepted task-configuration record is available.
This stage only; model and reasoning effort are not inferred from other stages. - Visual review
- Unknown · Reasoning effort:Unknown
No independently verifiable accepted task-configuration record is available.
This stage only; model and reasoning effort are not inferred from other stages. - Technical review
- Unknown · Reasoning effort:Unknown
No independently verifiable accepted task-configuration record is available.
This stage only; model and reasoning effort are not inferred from other stages.
Version history
- v1 ·
Preserve the original low-sample animation and add actual temporal denoising, a sanitized public native project, technical evidence and bilingual presentation.
- v2 ·
Add synchronized before/after video, dcc-media-enhance attribution and timeline diagnostics, retaining the browser-playback and continuous-motion acceptance limits.
- v3 ·
Add the complete refined 960 × 720 guided spatial-OIDN candidate and actual 144-frame technical evidence. Static contact sheets are reviewed and historical atadenoise remains separate. Add full actual-picture playback gates; current-head browser acceptance remains incomplete; the user approved the current film for publication via voice.
- v4 ·
Record complete natural-speed playback, actual pictures and Stop release for English/Chinese desktop/touch on tested commit 5080041. Touch end-picture inspection uses real native fullscreen controls; normal inline-control occlusion is retained. Later commits require new verification.
English translation. Read the original Chinese text. Source evidence, downloads and native artwork retain their original language.
Prompt
Keep the original brass hoop, porcelain pendulum and deep-teal plinth. Use genuine DCC-MCP edits to refine native geometry, damped swing and the 42-to-44 mm lens. Save and reopen Blender, then render mixed-precision multilayer linear EXR with Cycles in-render denoising disabled. Use the unified CLI and actual Albedo/Normal guides for independent per-frame spatial OIDN, followed by explicit AgX and sRGB-to-BT.709 delivery. Preserve originals and chain hashes; separately check exact timing, consecutive frames and full natural-speed playback. Clean local paths and metadata before publication.
Reusable instructions derived from the actual production process, not a verbatim user request or private task prompt.
Software and tools
- Blender
- Blender 4.5.13 LTS · Cycles CPU · 96 samples
- dcc-media-enhance
- 0.2.0 · bd4aca72488ca6165816ebdf95c99df5800a192a
- Actual denoising backend
- Separate Blender compositor OIDN; embedded OIDN version not independently verified
- EXR precision
- Combined / Albedo HALF16; Normal / Depth FLOAT32
- FFmpeg
- FFmpeg 7.1.5 · Explicit sRGB-to-BT.709 · H.264/yuv420p · Limited range
- DCC-MCP Core
- 0.20.44
- Adapter record
- dcc-mcp-blender 0.2.20 baseline and 397d2aaa camera/timebase source patch; frozen snapshot, not a new release
- Output and timebase
- 960 × 720 · 144 frames · 24000/1001 fps · 6.006 s
- Public native copy
- Paths cleaned, saved and reopened; equivalence is limited to inspected settings and all-144-frame bounds
- Browser acceptance environment
- Chromium 153.0.8010.12 · English/Chinese desktop and touch-viewport emulation
Blender DCC-MCPGenuine host calls edit native geometry, camera and scene settings, followed by save/reopen inspection.
Cycles CPU96 samples, fixed seed 42, no adaptive sampling or animated seed, and in-render denoising disabled; output real multilayer EXR and guide passes.
dcc-media-enhance 0.2.0 / Blender compositor OIDNThe unified native CLI uses a separate empty Blender process for per-frame spatial denoising with real Albedo/Normal guides; no temporal algorithm or motion compensation.
Blender AgX / FFmpeg zscaleFloat32 linear beauty passes through AgX to 16-bit PNG, then explicit sRGB-to-BT.709 encoding; no interpolation, retiming or upscaling.
Adapters and related projects
Refer to project documentation for installation, compatibility and current capabilities. This case's actual environment and execution evidence are recorded above.
Step-by-step process
Refine geometry and preserve full framing
26 objects; hoop control points 96 to 384, hoop bevel resolution 3 to 6 and rod bevel resolution 6. The 42-to-44 mm camera has a recorded minimum 4.73% margin across all 144 projected-boundary checks.
Preserve native linear input
Render 96-sample mixed-precision multilayer EXR after saving and reopening. Preserve original bytes; do not publish internal paths or raw execution records.
Actual guided spatial denoising
After the production host exits, a separate process invokes installed Blender compositor OIDN. Use real Albedo/Normal guides per frame and output float32 linear RGBA EXR.
Explicit color chain and separate playback acceptance
Create display PNG with AgX/Look None/exposure 0/gamma 1, then explicitly convert sRGB to BT.709. Establish chronology using actual EXR headers, chain hashes, conversion receipts and MP4 PTS; PNG Frame=001 metadata from the denoise-only scene is not timing evidence.
Results
Verification and boundaries
| Check | Result | Observations and supporting evidence |
|---|---|---|
| Public native copy save/reopen | Passed | 26 objects, no external links or embedded text scripts; inspected render settings and all 144 projected bounds match. No rerender established pixel identity. |
| Complete 144-frame technical chain | Passed | The complete 144-frame chain of original EXR, spatial-OIDN output, display PNG and delivery hashes is verified. The full film decodes with PTS increments of 1001, timebase 1/24000 and duration 6.006 seconds, without exact consecutive duplicate frames. |
| Whole-film static contact-sheet review | Passed | All six contact sheets covering 144 frames and native-size frames 1, 13, 73 and 144 were inspected. No blocking static defect was observed at those scales and the full plinth remains visible. Fine bright speckling or aliasing remains on ring/rod highlights and a few bob-edge pixels; this is not natural-speed motion acceptance. |
| Actual natural-speed picture review | Passed | Immutable tested commit 508004143e41e404b38b149a8baf05fd18ef7ece: all English/Chinese desktop/touch contexts played naturally at 1× through ended, decoding 144 frames with zero drops/corruption/seeks, 143 callbacks covering frames 2–144 and 142 visible changes. First/middle/final actual decoded pictures match delivery references. This establishes display consistency, not temporal-denoising quality. |
| Full browser playback on the recorded commit | Passed | All four Chromium 153.0.8010.12 contexts passed; see ci-verification.json. Touch playback stays inline, where normal native controls obscure part of the ended image and its pixel mismatch is retained. The same ended picture passes separately through real native enter/exit fullscreen controls, with time/decode state unchanged and Stop releasing media without refetch. Later documentation or code commits do not inherit this result. |
| High-sample ground truth and quantitative quality score | not_run | No higher-sample ground truth, PSNR/SSIM or objective perceptual-quality certification; pixel matching establishes consistency with the delivery reference only. |
| User acceptance of the current film | Passed | The current 6.006-second film has subjective user acceptance; this is not objective quality certification. Browser display consistency is separately bound to the explicitly recorded tested commit. |
Evidence limits
- This is an original procedural animation, not CineRecon reconstruction, a physical pendulum simulation, a benchmark or a model-quality score.
- OIDN runs independently per frame here and is spatial; there is no temporal algorithm or motion compensation, and still-frame success does not prove motion stability.
- The historical atadenoise experiment differs in geometry, camera, dimensions and sampling; its old raw video is not a same-source control for the new film.
- MP4 is a lossy display encode. Browser references are independently decoded from that final MP4 and verify display consistency, not native EXR or high-sample quality ground truth.
- The public native copy differs from production bytes; equivalence is limited to inspected settings and projected bounds, without pixel-identical rerender proof.
- Touch is Chromium viewport emulation. Normal inline native controls can obscure the ended image; its separate real-fullscreen inspection does not claim unobstructed inline pixels. Physical devices, Safari/Firefox and production hosting were not tested.
- No independently verifiable accepted task-configuration record establishes model or reasoning effort; every stage remains Unknown.
- dcc-media-enhance source is MIT-licensed; original case assets remain rights-reserved, and upstream Blender, OIDN and FFmpeg licensing obligations remain unchanged.
Last reviewed: 2026-10-16
Project files
Choose a project, reproduction package or exported resource. Before use, review this case's software versions, dependencies, format boundaries and rights.
- New candidate native project (paths cleaned)BLEND · 1.5 MiBDownload project
Resources and rights
- Bilingual case report
- Current native-edit recipe
- Current native inspection and cleanup scope
- Current technical verification
- Actual spatial-denoise and color-chain evidence
- Current visual review and unaccepted checks
- Frame order and exact timebase
- Actual software versions and capability boundaries
- Browser same-delivery reference contract
- Compact browser receipt for the tested commit
- File inventory and hashes
- Complete historical atadenoise case and old CI evidence
- Rights and usage boundaries
- Actual unified CLI source (MIT)
Source:Source and code
Source commit:bd4aca72488ca6165816ebdf95c99df5800a192a
License scope: Original scene, imagery, recipe and case materials: rights reserved. Public display grants no additional MIT license. Upstream software licenses are unchanged.
No external image textures, generative bitmap assets or soundtrack; all imagery originates in the native Blender scene.