Why Output Frames Do Not Define a Video Generation Workload

Start with the video the offer must deliver

A quote for one clip leaves much of a video-generation workload unspecified. Even adding an output frame count leaves the spatial dimensions unresolved. A second quote can repeat those labels while asking its model to process a much larger representation. Keeping the same model checkpoint does not remove that difference.

This guide examines the Wan2.1 text-to-video architecture in the Wan Team's March 26, 2025 report, arXiv:2503.20314v1, and the WanPipeline implementation tagged Diffusers v0.35.1. It explains a workload mechanism, without recommending a current model or transferring the result to other Wan pipelines.

The report describes a causal video variational autoencoder, Wan-VAE. For a video with 1 + T frames, height H and width W, its latent geometry is 1 + T/4 temporal positions by H/8 by W/8, with 16 channels. The first frame receives spatial compression only. That separate first position explains why dividing the total frame count by four gives the wrong temporal count.

References: [VID1]

Follow text, latent noise and decoded output separately

For the ordinary text-to-video path with no supplied latents, the tagged implementation encodes the prompt and initializes a noisy latent tensor. It iterates through scheduler timesteps, predicts from the current latent state and updates it. For frame output, it then applies the VAE decoder and video postprocessing.

The VAE's encoding relation explains the correspondence between video and latent geometry. Generated output frames are not first encoded as inputs to this text-to-video generation path. They are the eventual decoded result.

That order matters when defining a timing boundary. Text encoding, iterative latent processing and final decoding contribute different work. Evidence covering only the denoising portion leaves the time to obtain the finished video unresolved. A result that returns latents also leaves a different amount of work completed from a result that returns decoded frames.

References: [VID2]

Wan-WL112: 81 frames become 21,504 video positions

Original arithmetic example Wan-WL112 stipulates one clip and a configuration supporting all three shapes compared here, using the report's VAE geometry and 1 x 2 x 2 patchifier. This support is an invented assumption, not evidence that an offered checkpoint accepts these inputs. Height and width are divisible by 16, and F - 1 is divisible by four. No resizing, padding or additional shape transformation is included.

Let F denote total output frames. The paper's T is therefore F - 1. At F = 81 and H = W = 512 pixels, the latent temporal count is 1 + (81 - 1)/4 = 21. Its spatial dimensions are 512/8 by 512/8 = 64 by 64. The latent grid has 21 x 64 x 64 positions, each with 16 channel values.

The described patchifier groups one temporal position and a 2 x 2 spatial region, projecting patches into feature vectors. Flattening produces a video sequence of length L = (1 + (F - 1)/4) x (H/16) x (W/16). Here L = 21 x 32 x 32 = 21,504.

The 16 channels contribute values within each patch; they do not multiply the number of sequence positions by 16. The projected feature width is a separate dimension. Nor does this single clip become 81 independent batch items. Its temporal representation participates in one video sequence, alongside separately represented text conditioning.

The result is 21 latent temporal positions, rather than 81/4 = 20.25. Diffusers v0.35.1 obtains the temporal count from the supplied VAE scale factor using integer division of F - 1, followed by the extra first position. Its scale factors come from the VAE configuration, so an actual offer needs that configuration identified. The tagged path normalizes incompatible frame requests before latent preparation; obtain the effective accepted output count as well as the requested count.

References: [VID1][VID2]

Adding frames changes the temporal axis

Keep Wan-WL112's 512 x 512-pixel dimensions and increase F to 161. The latent temporal count becomes 1 + (161 - 1)/4 = 41. The grid is now 41 x 64 x 64, and patchification gives 41 x 32 x 32 = 41,984 video sequence positions.

The sequence ratio is 41,984/21,504 = 41/21, approximately 1.952. The output frame ratio is 161/81, approximately 1.988. They differ because both cases retain the separate first temporal position.

This is a longer representation of one clip. It supplies no evidence that temporal slices can be scheduled as independent requests or that twice the output duration produces exactly twice the execution time. For a buyer changing clip length, the useful new information is the changed latent time axis and resulting sequence presented to the model.

References: [VID1]

The same 81 frames can produce four times the sequence

Return to F = 81, but increase both height and width from 512 to 1,024 pixels. Wan-WL112 now has a 21 x 128 x 128 latent grid. Its patchified video sequence contains 21 x 64 x 64 = 86,016 positions.

The temporal count is unchanged. Each spatial axis doubles, so their product quadruples: 86,016/21,504 = 4. Two offers described as one 81-frame clip can consequently describe fourfold-different video sequence populations under these assumptions.

This ratio concerns sequence positions. It does not establish four times total GPU memory, latency or price. A conceptual dense attention relationship also does not prove that an implementation allocates a complete square attention array. The related FlashAttention guide explains that allocation boundary; the allocator guide explains why tensor, managed-pool and whole-device memory reports need separate interpretation.

The model weights can remain the same while the live latent and activation shapes change. A weight-size figure describes the parameters being loaded; it cannot settle the memory requirement of these different active shapes.

References: [VID1]

A denoising iteration updates the representation

In the tagged Wan2.1 path, a transformer prediction uses the current latent state, timestep and text conditioning; the scheduler uses the prediction to update the state. Classifier-free guidance can add a separate prediction using negative-prompt conditioning before the update.

A loop iteration therefore concerns an update across the video's latent representation. It does not mean that one more output frame has been decoded. Likewise, the requested inference-step count, actual scheduler timestep sequence and number of transformer evaluations need their own definitions. The guidance branch already prevents a universal one-model-call-per-step rule.

For Wan-WL112, increasing the sequence population and increasing the number of evaluations are different changes. One changes the representation processed by an evaluation; the other changes how often prediction work is performed. Matching frame counts establishes neither. Matching a nominal step count still leaves the scheduler and guidance path unresolved.

Ask the software provider to identify those conditions with the exact pipeline and implementation revisions. Fewer evaluations alone establish no equivalent output quality. Any comparison needs the application's accepted result kept visible while the execution conditions are reconciled.

References: [VID2]

Compare delivered video under a common requirement

The Diffusers v0.35.1 guide passes 81 frames to generation and separately supplies 16 frames per second when exporting the result. Its output definition keeps batch, frame, channel, height and width as distinct axes. The export cadence describes playback; it supplies no measurement of generated frames per wall-clock second.

Consider fictional offers A and B using Wan-WL112's supported-shape assumptions and the same checkpoint. Both advertise one 81-frame clip. A specifies 512 x 512 pixels; B specifies 1,024 x 1,024. Their 21,504 and 86,016 video positions expose a missing comparison condition. Neither offer has a stipulated runtime, peak-memory reading, price or quality result.

First resolve the application's required dimensions, clip length and playback cadence, together with acceptable detail, motion continuity and prompt adherence. The larger shape alone proves none of those quality outcomes. If both offers can deliver the agreed output, obtain evidence for that common requirement. If they deliver different outputs, preserve that difference instead of treating the shared clip label as equivalence.

The remaining evidence should connect the supported output shape to the actual latent and patch configuration, scheduler, guidance path and executed implementation. Keep precision, batch size and concurrent clip requests attached to the result. Ask for peak-memory coverage and units across the relevant phases, and elapsed time through decoded output. Identify whether queueing, prompt processing and video export are inside the reported completion interval.

A price per clip then needs the same accepted deliverable and included work. A provider's metering definition is separate from the architecture's sequence count. Use the related benchmark guide for the broader system and measurement comparison; the missing evidence here is specifically how this video is represented and repeatedly processed.

The raw-image ViT guide addresses a different conversion: image pixels into a classification encoder's patch sequence. Wan-WL112 adds a temporal VAE relation, noisy latent evolution and decoding. Multiplying an image-patch example by output frames would miss those stages.

References: [VID3]

Scope a separate physical equipment requirement

If the selected system also needs equipment receiving, rack integration, onsite installation or IT cabling, Leviathan's Complete Physical Infrastructure Deployment service is the relevant next step. Use the related service and project inquiry with a nonconfidential equipment list, site, required physical stages and handoff needs to discuss scope and availability.

Model serving, inference evaluation, tuning and application operation remain with the appointed software team. The physical inquiry concerns installation of separately selected equipment.

References: [LS1]