Local Image-to-Video Generation Hardware: VRAM, Generation Time, and Realistic Minimums
Documented VRAM minimums are not the real number. Here is what each open-weights image-to-video model actually needs with the full ComfyUI pipeline, and why the gap exists.
Local Image-to-Video Generation Hardware: VRAM, Generation Time, and Realistic Minimums
You have a workflow loaded in your image-to-video tool, a source image cropped to 1024x576, and the Wan 2.1 14B model selected. You press Queue and wait. Ten seconds later, your RTX 3060 returns an out-of-memory error and the console logs a CUDA allocation failure. The model card said 24 GB. You have 12 GB, but you expected quantisation to bridge the gap. It did not. This is the central truth of image-to-video local generation hardware VRAM requirements: the documented minimum almost never matches what you actually need when the full pipeline runs.
The gap between claimed VRAM and real peak usage is this page's subject. For every open-weights model, Stable Video Diffusion, LTX-Video, and Wan 2.1, you will find the manufacturer's documented minimum and then the realistic minimum for a full run that includes VAE decode and resolution enhancement. The difference is 2-4 GB. If you are on an 8 GB card, that gap is the difference between a generation and a crash. Here are the numbers, the structural trade-offs, and the tier of hardware each model actually demands so you can decide which GPU to buy or whether your existing card will work.
Stable Video Diffusion Minimum VRAM Consumer GPU: Claimed vs. Real
Stable Video Diffusion (SVD) lists a minimum VRAM of 16 GB on its model card dated 2023-11-21. That figure reflects inference at batch size 1 in fp16 precision, generating the native 25 frames at 1024x576 resolution. No VAE decode, no resolution enhancement, no tooling overhead. In practice, a full pipeline, loading the ImageOnlyCheckpointLoader, decoding the latent output to pixel space, and saving as MP4, pushes peak VRAM to 12-14 GB. That is still under the claimed minimum, but only because the 16 GB figure is generous for the base case. The real problem starts when you add any post-processing: enhancing to 1080p or running frame interpolation adds 1-2 GB on top, putting you at 14-16 GB total. An 8 GB card will not run SVD at all. A 12 GB card runs it but leaves no room for the tool's own memory allocation, so expect random OOM errors on longer workflows.
Generation time on an RTX 4090 is 10-15 seconds for the 25-frame clip. On an Apple M2 Max with 32 GB unified memory, expect 60-90 seconds per generation. The 25-frame clip at 24 fps produces just over one second of video. If you need a longer clip, you must concatenate multiple generations, which introduces seam artefacts, visible jumps at the join, because SVD does not accept an end frame to condition the transition. The motion bucket ID parameter (range 1-255, default 127) controls how much motion the model applies; higher values increase motion strength but also increase the risk of subject drift and limb mutation. The augmentation level parameter (0.0-1.0) adds noise variation, but raising it above 0.2 degrades source-image fidelity.
LTX-Video Consumer GPU Inference Time: Fast on 8 GB, But Short
LTX-Video from Lightricks is the fastest open-weights image-to-video model on consumer hardware. Its model card (dated 2024-12-05) states a minimum of 8 GB VRAM, and that figure is realistic for the bf16 precision it uses by default. Peak VRAM usage for a 121-frame clip at 768x512 resolution runs 7-9 GB during inference, including VAE decode. That puts it within reach of an RTX 3060 12 GB or an RTX 4060 Ti 16 GB. An 8 GB card like the RTX 3070 runs it, but only just: expect OOM errors if the tool loads any additional nodes for resolution enhancement or frame interpolation. The recommended VRAM of 12 GB gives comfortable headroom for the full pipeline and allows switching to fp16 precision for slightly better quality.
Inference time on an RTX 4090 is 15-30 seconds for the full 121 frames (5.04 moments of video at 24 fps). On an Apple M2 Max 32 GB, expect 2-4 minutes. The guidance scale (default 3.0, range 1.0-10.0) controls how closely the output follows the source image; values above 6.0 produce oversaturated colours and motion collapse, where the clip becomes nearly static. The num_inference_steps parameter (default 40, range 20-50) trades quality for speed: at 20 steps, generation time on an RTX 4090 drops to under 10 seconds, but subject drift and flicker increase noticeably. LTX-Video supports end-frame conditioning via its i2v-end-frame workflow, which lets you specify the last frame of the clip. That reduces seam artefacts when concatenating clips, but you need to generate the end frame separately first, which doubles total generation time.
| Model | Documented Minimum VRAM | Realistic Minimum VRAM (Full Pipeline) | Generation Time (RTX 4090) | Native Clip Length (Seconds) |
|---|---|---|---|---|
| Stable Video Diffusion | 16 GB | 14–16 GB (with upscaling) | 10–15 s (25 frames) | 1.04 s at 24 fps |
| Wan 2.1 1.3B | 8 GB | 6–8 GB | 30–60 s (81 frames) | 5.06 s at 16 fps |
| Wan 2.1 14B | 24 GB | 20–22 GB | 5–10 min (81 frames) | 5.06 s at 16 fps |
| LTX-Video | 8 GB | 7–9 GB | 15–30 s (121 frames) | 5.04 s at 24 fps |
Wan 2.1 14B VRAM Requirement: The 24 GB Cliff
Wan 2.1 from Alibaba comes in two parameter sizes. The 1.3B model has a documented minimum of 8 GB VRAM and peaks at 6-8 GB during a full run generating 81 frames at up to 1280x720. That makes it the only open-weights model that genuinely runs on an 8 GB card without OOM errors. Generation time on an RTX 4090 is 30-60 seconds; on an Apple M2 Max 32 GB, expect 3-5 minutes. The 1.3B model uses fp16 precision and is Apache 2.0 licensed, so commercial use is permitted. The motion_scale parameter (range 0.0-1.0) controls motion magnitude; values above 0.7 produce rapid subject drift and limb mutation, especially on subjects with fine details like hands or hair.
Wan 14B Hardware Requirements
The 14B model is a different class of hardware requirement entirely. Its model card (dated 2025-02-26) states a minimum of 24 GB VRAM. Realistic peak usage in a full pipeline, including VAE decode, resolution enhancement to 1280x720, and saving as MP4, runs 20-22 GB. That is within 2 GB of the documented figure, but the gap matters because the tool's own memory allocation adds overhead. On a 24 GB card like an RTX 4090, the full pipeline works with no headroom for additional nodes. A 32 GB card (RTX 6000, A6000) gives comfortable space for enhancement nodes, frame interpolation, or running multiple batch sizes. Generation time on an RTX 4090 is 5-10 minutes for the 81-frame clip (5.06 moments of video at 16 fps). The 14B model uses bf16 precision in its default workflow; switching to fp16 is possible on 32 GB cards but offers negligible quality gain for a 10% VRAM increase.
Why the Gap Matters More for Wan 14B than SVD or LTX
SVD's documented minimum is 16 GB and its real peak is 14-16 GB, a range that fits on a 24 GB card comfortably. LTX-Video's 8 GB minimum and 7-9 GB real peak fit on any card sold in the last five years. Wan 14B's 24 GB minimum and 20-22 GB real peak put it right at the boundary of what a single RTX 4090 can deliver. Any memory leak in a custom node, any attempt to enhance to 1080p inside the same pipeline, or any batch size above 1 will push it to 24 GB and crash it. If you are on an RTX 3090 (24 GB), run Wan 14B in isolation with no other applications open and no browser tabs rendering GPU-accelerated content. If you are on an RTX 3080 Ti (12 GB), skip Wan 14B entirely and use the 1.3B model or LTX-Video.
Image-to-Video Documented vs. Actual VRAM Usage Gap: Why It Exists and What to Do About It
The gap between documented and actual VRAM usage exists because model-card authors optimise for a clean test environment: a single inference pass with no post-processing, no VAE decode, and no node overhead. The real world adds these layers. The VAE decode step alone can consume 1-3 GB depending on the latent size and the VAE model used. Enhancing from native resolution (1024x576 for SVD, 768x512 for LTX, 480x480 for Wan 1.3B) to 1080p adds another 1-2 GB. Frame interpolation nodes add 500 MB to 1 GB. The runtime itself allocates around 500 MB of VRAM for its graph execution engine. Sum these, and the documented minimum is always at least 2 GB lower than the real peak for any model that requires VAE decode.
Quantisation Trade-offs
Quantisation helps but does not eliminate the gap. Using fp8 quantisation on SVD or Wan reduces VRAM by approximately 30%, but the quality loss, especially in fine detail like facial features and text, is noticeable on any output above 720p. The practical trade-off is: use fp8 only if you are VRAM-constrained and generating short clips (under 3 moments) where quality degradation is less visible. For longer clips, fp16 or bf16 is mandatory to maintain temporal consistency. Driver updates also shift these numbers: CUDA 12.1, which the tooling recommends, allocates memory more efficiently than 11.8, reducing peak usage by 500 MB to 1 GB on some models. If you are on an RTX 30-series card, update to CUDA 12.1 before blaming your GPU.
What You Actually Get at Each Tier
The durable takeaway is that no open-weights image-to-video model does long, high-resolution, and fast on 8 GB. LTX-Video gives you 5 moments of 768x512 video in under 30 seconds on an RTX 4090. Wan 2.1 1.3B gives you 5 moments of 720p video in 30-60 seconds on the same card. Wan 2.1 14B gives you 5 moments of 720p video in 5-10 minutes on a 24 GB card. None of these produce 10-moment clips at 1080p without enhancement and concatenation, which add generation time and introduce seam artefacts. If you need longer clips, accept lower resolution or longer generation times as the cost of the hardware you have.
Realistic Minimums: What Each Model Tier Actually Needs
The structural relationship between model parameter size and hardware requirement is simple: parameter count drives VRAM, and VRAM drives cost. Wan 2.1 1.3B (2.5 GB on disk) and LTX-Video (2 GB on disk) both run on 8 GB cards, but only LTX-Video gives you 5 moments of video in under 30 seconds. Wan 1.3B is slower, 30-60 seconds for the same clip length, and produces softer output at its native 480x480 resolution. Enhancing to 720p is necessary for most use cases, which adds 1-2 moments of FFmpeg post-processing time but does not increase VRAM.
Choosing a GPU by Model Tier
Stable Video Diffusion (5 GB on disk) requires 16 GB VRAM for a full pipeline. That means an RTX 3080 12 GB is insufficient. An RTX 3090 24 GB works, but the 25-frame native clip length (1 moment) limits its practical use to establishing shots or transitions. The 1-moment clip is too short for most video projects; concatenating multiple SVD clips creates visible seam artefacts because SVD lacks end-frame conditioning. LTX-Video's 121-frame clip (5 moments) and Wan 2.1's 81-frame clip (5 moments) are the minimum practical clip lengths for b-roll or social media video. Wan 2.1 14B (28 GB on disk) demands a 24 GB card and delivers the highest motion coherence and source-image fidelity of the three models at 1280x720. Its generation time of 5-10 minutes on an RTX 4090 makes it unsuitable for interactive workflows but acceptable for batch generation.
The Hard Floors
The most common failure case is buying an RTX 3060 12 GB expecting it to run Wan 14B after quantisation. It will not. The data is clear: 24 GB is a hard floor for Wan 14B, and 8 GB is a soft floor for LTX-Video and Wan 1.3B. If you are on an 8 GB card, run LTX-Video in bf16 precision with no enhancement nodes and expect 15-30 second generation times. If you are on a 12 GB card, run Wan 1.3B or LTX-Video but skip Wan 14B entirely. If you are on a 24 GB card, you can run all three models, but Wan 14B will dominate your GPU for 5-10 minutes per generation. Check your GPU's VRAM against the realistic minimums here before downloading any model weights. The model card says one thing; the full pipeline says another. Trust the pipeline.
Common Questions
Can I run image-to-video models on an AMD or Intel GPU?
No. NVIDIA CUDA is required for Stable Video Diffusion, Wan 2.1, and LTX-Video. Apple M-series GPUs work via the MPS backend, but only LTX-Video and Wan 1.3B have tested MPS support. AMD ROCm support is not available for any of these models.
Why does my 8 GB card OOM even with LTX-Video, which says 8 GB minimum?
The documented 8 GB is for batch-size-1 inference only. If the tool loads any additional nodes for resolution enhancement, frame interpolation, or VAE decode, peak usage can hit 9 GB. Close all other GPU-using applications and run the workflow in isolation.
How much system RAM do I need for local image-to-video generation?
The recommended system RAM is 64 GB. Minimum 32 GB. System RAM is used for loading model weights before they are offloaded to VRAM and for the graph execution cache. Below 32 GB, you will see disk swapping that multiplies generation time by 3-5x.
Does quantisation (fp8) reliably reduce VRAM by 30% on a 24 GB card?
fp8 quantisation reduces VRAM by approximately 30% on SVD and Wan 14B, but quality loss in facial features and text is noticeable at 1080p output. Test a single generation before committing to a workflow batch. The reduction is not guaranteed on all custom nodes.