For a separate game-rendering release focused on entity correspondence, use the AnyGameRenderer input and review guide.
What SolarRenderer does
SolarRenderer is a structure-conditioned video renderer presented by its authors in an October 2026 release. The paper page describes adapting a frozen MiniMax-H3 backbone with a control LoRA and a few-step LoRA. Its production question is useful for AI cinema: can you establish a scene's movement and camera in a structural clip before asking a renderer to supply appearance?
This guide separates that workflow from ordinary text-to-video prompting. It includes our own downloadable greyscale source, frame-zero image and review brief. We checked the public project, model card and documentation. We did not install SolarRenderer, download its weights, submit the source to a model or measure an inference result.
Inspect our original structure input
Watch this clip and view its details
Open the standalone original study for the video and its source guide.
One greyscale capsule figure travels from left to right behind a fixed central column. The camera stays fixed. The input is a code-authored two-dimensional blockout with simple shading, not a recorded actor, a depth map or a SolarRenderer result. Its purpose is to make the intended trajectory and temporary occlusion inspectable.
The MP4 has 158 frames at 24 fps and a 1344×768 picture. Its exact frame-based duration is 158 divided by 24, approximately 6.583 seconds. These measured source properties do not prove acceptance by every runtime. The reference inference guide specifies a 24 fps structure clip with at least 158 frames for the default landscape or portrait canvas. Check the current route before using the file.
Keep structure appearance and review separate
| Material | What it should communicate | Our supplied file |
|---|---|---|
| Structure video | Trajectory camera and occluding column | Original greyscale MP4 |
| Frame zero | The source composition before the crossing | Extracted source PNG |
| Desired appearance | Materials lighting and fictional subject design | A creative brief to develop separately |
| Accepted result | What the actual complete render preserved | No result supplied or claimed |
Our frame-zero PNG is still greyscale. It is not a finished style reference. If you develop a styled first frame, preserve the opening layout and subject position rather than inventing a different camera. Keep that image and the source side by side so you can see what was changed intentionally.
Choose the current runtime and matching adapter pair
The official repository lists a reference script, a ComfyUI route, a browser demo and a community Reactor integration. These are distinct entry points. Save the exact checkpoint, adapter pair, workflow and version for the route you actually use. Do not treat a screenshot of one interface as proof that a different package accepts identical settings.
The weight card distinguishes full-clip control and four-step adapters from their causal counterparts. Keep the intended pair together. Its code and weights use different licenses; the weight card lists territorial exclusions including the United States, European Union, United Kingdom and Republic of Korea. Read the linked weight terms before use. Availability of a download is not a blanket usage grant.
For the ComfyUI route, use its current component list and workflow rather than mixing model formats from a different inference package. The documentation supplies full-quality, low-memory and longer-clip workflows. Its hardware guidance is route-specific, not a guarantee established on our machine.
Write appearance into the documented prompt form
The official prompt guide uses six sections: subject definitions, summary, retention analysis, detailed description, overall soundscape and non-diegetic music. It describes a 350–500-word detailed-description section and an assembly and validation process. Follow that current process for the selected condition rather than treating our short creative brief as a ready-to-run prompt.
For our Column Crossing exercise, describe one fictional adult-sized courier form, a fixed archive camera and the same left-to-right crossing. Let the appearance treatment add an olive fabric coat, pale archive plaster and soft overhead light without adding a second traveler. Keep the column in front of the traveler during the middle portion. Those are intended review targets, not claims about what a submitted model run achieved.
The source has no soundtrack. If a future run requests footsteps or ambience, record that as a separate requested property and inspect the returned audio. Do not infer spoken words, synchronized sound or a specific audio track from the silent structure file.
Review geometry and occlusion before judging style
Compare source and returned video at matching times. Check where the figure starts, where it disappears behind the column and where it reappears. Look for duplicated figures, motion reversals, a column that becomes transparent or a camera that moves despite the fixed plan. These defects matter even if an individual output frame looks attractive.
Then review appearance across the whole clip. The figure should retain its chosen wardrobe and shape after reappearance, while the room and light should remain coherent. Keep the original source and accepted first frame available. A second run should change a specific failed variable rather than silently replacing the entire brief.
Separate release claims from your measured turnaround
The project page reports four-step rendering and a block-causal streaming variant. Those descriptions concern the authors' methods and demonstrations. We have not reproduced their speed, streaming latency or benchmark scores. An offline clip, a queued browser run and a streaming service have different timing boundaries.
Record preparation, queue, sampling, transfer and finishing where they affect the delivery decision. A fast returned file that loses the traveler behind the column permanently is a rejected take. Use actual charges and elapsed time for your route; no local cost or accepted-output rate is established by our input fixture.
Download the input and pilot record
Download the original structure MP4, source frame zero, Column Crossing creative brief, blank review ledger and scoped file report. Result and acceptance fields remain empty because no inference was performed.
For original cast and scene development, explore Clout AI Cinema and its currently supported operations. This guide does not establish a SolarRenderer integration in Clout. The previsualization workflow covers broader scene planning; the FastH3 Omni guide covers a separate reference release and runtime decision. Keep their input contracts distinct.
