import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("backpack-run/Wan2.2-TI2V-5B-Backpack-Video", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]Wan2.2 TI2V-5B — Backpack Pipeline Package
Public Apache-2.0 pipeline supporting text-to-video and image-to-video at 720p. This hybrid package contains a material Backpack-produced model component and resolves unchanged components from immutable upstream revisions.
Pipeline
| Field | Value |
|---|---|
| Capability | video-generation |
| Tasks | text-to-video, image-to-video |
| Input | text, image |
| Output | video |
| Runtime | diffusers 0.40.0 |
| Pipeline class | WanPipeline |
| Artifact strategy | backpack-artifact |
Components
| Component | Class | Immutable source | Size |
|---|---|---|---|
| transformer | WanTransformer3DModel |
Wan-AI/Wan2.2-TI2V-5B-Diffusers@b8fff7315c768468a5333511427288870b2e9635 |
18.6 GiB |
| text_encoder | UMT5EncoderModel |
Wan-AI/Wan2.2-TI2V-5B-Diffusers@b8fff7315c768468a5333511427288870b2e9635 |
10.6 GiB |
| tokenizer | T5TokenizerFast |
Wan-AI/Wan2.2-TI2V-5B-Diffusers@b8fff7315c768468a5333511427288870b2e9635 |
20.5 MiB |
| vae | AutoencoderKLWan |
Wan-AI/Wan2.2-TI2V-5B-Diffusers@b8fff7315c768468a5333511427288870b2e9635 |
2.6 GiB |
| scheduler | UniPCMultistepScheduler |
Wan-AI/Wan2.2-TI2V-5B-Diffusers@b8fff7315c768468a5333511427288870b2e9635 |
820.0 B |
| pipeline | WanPipeline |
Wan-AI/Wan2.2-TI2V-5B-Diffusers@b8fff7315c768468a5333511427288870b2e9635 |
499.0 B |
Execution profiles
reference
BF16 reference weights with documented component offload and FP32 VAE decode.
| Component | Precision | Device | Quantization |
|---|---|---|---|
| transformer | bf16 | cuda | none |
| text_encoder | bf16 | cpu | none |
| vae | fp32 | cuda | none |
nvidia-fp8
TorchAO FP8 weight-only transformer with Diffusers model CPU offload.
| Component | Precision | Device | Quantization |
|---|---|---|---|
| transformer | fp8 | cuda | torchao / Float8WeightOnlyConfig |
| text_encoder | bf16 | cpu | none |
| vae | fp32 | cuda | none |
compatible-int8
TorchAO INT8 weight-only transformer with Diffusers model CPU offload.
| Component | Precision | Device | Quantization |
|---|---|---|---|
| transformer | int8 | cuda | torchao / Int8WeightOnlyConfig |
| text_encoder | bf16 | cpu | none |
| vae | fp32 | cuda | none |
Backpack-produced artifacts
| Artifact | Component | Profile | Format | Precision | Size | Load check |
|---|---|---|---|---:|---|
| transformer-compatible-int8 | transformer | compatible-int8 | diffusers-torchao-safetensors | int8 | 4.7 GiB | passed |
Unchanged text encoder, tokenizer, VAE, scheduler, and pipeline configuration remain pinned upstream and are not duplicated merely to make the package self-contained.
Hardware
The native upstream documents at least 24 GB VRAM for its single-GPU offloaded command; optimized Diffusers profiles require measurement. These values are marked
upstream-reported; optimized profiles still require measured
peak-memory and timing reports.
Referenced component size is 31.9 GiB.
Recommended inference workspace is
36.0 GiB
(calculated): Workspace estimate is upstream component bytes x 1.1 plus 1 GiB; it excludes generated outputs and duplicated caches.
Validation
Static validation passed for the immutable revisions, Apache-2.0 license metadata, public access, component inventory, pinned runtime, execution-profile structure, and remote integrity values. The transformed component passed serialization and local reload validation, but full pipeline inference was not executed. Prompt adherence and perceptual quality have not been reviewed. This materially transformed artifact may be published after the required package checks pass; inference/output validation remains pending and is reported separately.
Provenance and integrity
- Upstream: Wan-AI/Wan2.2-TI2V-5B
- Immutable upstream revision:
921dbaf3f1674a56f47e83fb80a34bac8a8f203e - Runtime artifact inventory:
upstream-artifacts.json - The inventory records SHA-256 for LFS objects and Git object SHA-1 for ordinary repository files.
- Backpack did not train this model and does not claim ownership of it.
- License and attribution files:
LICENSE,NOTICE
Review the upstream model card and license before use. Generated media may contain factual, representational, safety, or quality defects.
- Downloads last month
- -
Model tree for backpack-run/Wan2.2-TI2V-5B-Backpack-Video
Base model
Wan-AI/Wan2.2-TI2V-5B