Wan 2.2 and MiniMax: choosing a local video workflow

If you want a five-second promotional clip that animates a product photo, first decide whether to generate sound as well. MiniMax H3-Base outputs video and stereo audio together. For a silent clip, a Wan video workflow can be followed by adding effects or music in an editor.

Joint output follows a different computation path from making silent video and adding music later. Check that the model's output fits your editing workflow; if video and audio are saved together, play them back to review timing and whether the sound fits the scene.

To preserve the shape in an existing bottle photo, choose a model and runtime that accept image input. Adding an image option to a text-only checkpoint² does not make it image-conditioned; check what inputs the actual model files and workflow accept.

The model and tool must both support specifying first and last frames³. Fixing those endpoints does not determine the motion between them, so if both opening and closing scenes matter, review a short draft to see whether the transition is natural.

In addition to weights, the input image, intermediate video-frame data, and resolution use working memory. Offloading⁴ moves some data to host memory and can make a smaller-VRAM setup possible, but it also needs system RAM⁵ and transfer time. Multi-GPU setups require model and runtime support for partitioning, so memory capacities cannot simply be added as if they provided the same execution path.

The official Wan 2.2 TI2V-5B repository covers a 720p path at 1280×704 using a 24 GB GPU⁶ with offloading. Even when starting with a smaller-weight configuration, play the full clip to check whether the required movement and product outline remain. Resolution and frame rules differ by model, so do not transfer one model's settings unchanged to another.

LTX-2.5 requires width and height to be multiples of 32 and frame count to follow 8n+1. This site's HunyuanVideo 1.5 timing comparison covers a 480p image-to-video distillation path. These examples describe each model's own input conditions; do not use them as Wan settings. First save a short result within the target model's documented requirements.

Wan 2.2 T2V-14B is a text-to-video path, so its goal differs from TI2V-5B, which starts from a bottle photo. It is an option when you want a new scene from text without a reference image; check how the requested camera movement and product-like shape appear in the result.

A larger model changes not only storage needs but also loading and generation memory. Keep the prompt⁷ and scene goal fixed when comparing results, then decide whether improved appearance or motion justifies the extra memory and wait.

LightX2V's MiniMax-H3 guide separates Turbo 4-step and 8-step settings by precision and GPU parallelism. The 4-step BF16⁸ path uses dmd/minimax_h3_bf16_4step.json and a single-GPU script; 4-step FP8/INT8 uses SP4, while 8-step FP8/INT8 uses SP8. The FP8⁹ 4-step entry marked for 5090 is also an SP8 configuration, not a one-GPU run.

For a first setup, prepare the model directory according to the repository's model paths and kernel requirements. The documented Diffusers layout includes transformer¹⁰, text_encoder, tokenizer, processor, vae¹¹, and audio_vae components, so downloading only the original FL2VA/ checkpoint directory is not enough. Use the command below to check output from the baseline BF16 path.

After confirming that the baseline output file opens, copy dmd/minimax_h3_bf16_4step.json for the experiment. Check that the LoRA¹² path points to a real file and that alpha=128, video_flow_shift=6, audio_flow_shift=3, and infer_steps=4 match the published 4-step adapter.

The default is 362 frames at 768×1344, so for the first comparison reduce num_frames and size in the JSON to 124 frames at 544×960. Choose either dynamic LoRA application or merging into BF16 weights.

Keep prompt, seed, resolution, and frame count fixed, and save the baseline and Turbo outputs separately. Record first-run time including model loading and warm-up, repeat generation after setup, and time through file saving; these describe the wait in an actual workflow. The README has no consumer-GPU FPS¹³ comparison of the two presets under matched conditions, so do not infer a speed multiplier from four steps.

If the image is black or audio is missing, check the requested task, the --model-variant fl2av setting at startup, the model-component directories, and the LoRA path. VAE and encoder errors in the logs help distinguish an output-path problem from a model-loading problem.

LightX2V's H3-Base timing uses the SGLang FL2VA BF16/FP32 TP2 offloading path, two RTX 5090 cards, and 384 GB of host RAM. Matching only the two GPUs does not reproduce the same timing; offloading and host memory are part of the setup.

The measured scene is a 768p, five-second output with audio. The reported runs use five and 50 steps; the 20-step time mentioned as a standard setting is interpolated between those measurements. It is not a separate measurement and should not be extrapolated to other clip lengths or RAM configurations.

The published timing excludes prompt interpretation and the 2K regeneration stage. Use it when deciding whether joint video-and-audio output is useful or whether you would rather add a track to a silent clip in an editor. Review visual quality and audio fit by playing the result.

If the bottle photo must be preserved, compare an image-conditioned TI2V path; for a new scene from text, consider T2V-14B; if video and sound must be created together, consider H3-Base joint output. This is a choice about matching input and desired result, not a size ranking.

Videos shown on the site are prepared samples that illustrate estimated wait-time differences by hardware, not outputs generated by the actual models. Check whether the runtime you plan to install supports the model, image input, and required duration and resolution; save a short clip and play it from beginning to end. For frequent revisions, include first-run time, memory headroom, and editing time as well as generation wait.

If the target clip plays through on your current hardware, includes the audio you need, and repeat waits are manageable, you can start with that setup. When considering new hardware, compare run reports that specify model, precision, resolution, duration, and offloading. Treat reports under different conditions as context, not as estimates for your workload.