Update TI2V-5B memory and validation card

README.md

@@ -9,6 +9,7 @@ tags:
 - mflux
 - apple-silicon
 - 8-bit
+- mixed-q8-bf16
 - wan
 - wan2.2
 - video-generation
@@ -17,37 +18,62 @@ tags:
 ---
 # wan2.2-ti2v-5b-diffusers-8bit
 
-This repository contains MLX-Gen saved weights for `Wan-AI/Wan2.2-TI2V-5B-Diffusers`. The checkpoint is designed for local Apple Silicon inference with [`mlx-gen`](https://github.com/lpalbou/mlx-gen).
+This repository contains mixed q8/BF16 MLX-Gen saved weights for
+[`Wan-AI/Wan2.2-TI2V-5B-Diffusers`](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B-Diffusers).
+It is designed for local Apple Silicon inference with
+[`mlx-gen`](https://github.com/lpalbou/mlx-gen).
 
-It uses the mflux/MLX saved-weight layout and MLX quantization tensors. It is not a Diffusers or Transformers `from_pretrained()` checkpoint.
+It uses the mflux/MLX saved-weight layout with MLX quantization tensors. It is not a Diffusers or
+Transformers `from_pretrained()` checkpoint.
 
 ## Source Model
 
 Original model: [`Wan-AI/Wan2.2-TI2V-5B-Diffusers`](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B-Diffusers).
 
-## License and Access
-
 This quantized derivative follows the Apache 2.0 license of the source model.
 
 ## Quantization
 
-This is an MLX q8 checkpoint for Wan2.2 TI2V-5B. MLX-Gen uses 8-bit quantization for Wan modules where MLX supports quantization:
+This is a mixed q8/BF16 checkpoint:
 
-- q8 for quantizable Wan transformer attention and feed-forward modules.
+- q8 for quantizable Wan transformer attention and feed-forward linears.
 - BF16 for the Wan VAE.
-- BF16 for Wan transformer conditioning/output projection linears, the UMT5 text encoder, scheduler metadata, tokenizer files, norms, convolutions, and other non-quantizable parameters.
+- BF16 for Wan transformer `condition_embedder.*` and `proj_out`.
+- BF16 for the UMT5 text encoder, scheduler metadata, tokenizer files, norms, convolutions, and
+  other non-quantizable parameters.
+
+The upstream TI2V-5B source snapshot is not uniformly 16-bit on disk: the transformer and VAE
+safetensors are FP32, while the UMT5 text encoder is BF16. MLX-Gen loads Wan transformer/VAE
+weights at BF16 runtime precision.
+
+## Measurements
 
-Wan q4 quality and any possible mixed q4/q8 policy are still under validation. Prefer q8 for publishable Wan checkpoints until the q4 policy is documented.
+Measured on 2026-06-04 with `mlx-gen 0.18.10` on an Apple M5 Max with 128 GiB unified memory.
 
-See the [MLX-Gen quantization docs](https://github.com/lpalbou/mlx-gen/blob/main/docs/quantization.md) for compatibility notes.
+Validation profile: `1280x704`, 17 frames, 20 denoising steps, guidance `5`, 24 fps, seed `321`,
+explicit empty negative prompt.
 
-## Compatibility
+| Layout | Storage | Logical Model | Full-Process Physical Peak | Max RSS | MLX Peak | Total Time | Output |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | --- |
+| Upstream source snapshot | 31.9 GiB | 10.6 GiB | 102.7 GiB | 13.7 GiB | 58.5 GiB | 216.2 s | [base-source.mp4](validation/ti2v5b-clean/base-source.mp4) |
+| Prepared BF16 package | 21.2 GiB | 10.6 GiB | 102.6 GiB | 14.5 GiB | 58.5 GiB | 261.6 s | [prepared-bf16.mp4](validation/ti2v5b-clean/prepared-bf16.mp4) |
+| This mixed q8/BF16 package | 16.9 GiB | 6.3 GiB | 103.7 GiB | 13.8 GiB | 54.2 GiB | 243.4 s | [mixed-q8-bf16.mp4](validation/ti2v5b-clean/mixed-q8-bf16.mp4) |
 
-Requires `mlx-gen >= 0.18.6`.
+This package reduces storage, logical model bytes, active MLX model bytes, and MLX allocator peak in
+the validation profile. It did not reduce full-process physical peak memory in this profile because
+transient video-generation allocations dominated the run.
 
-Generated with `mlx-gen 0.18.6`.
+The source and prepared BF16 package produced byte-identical decoded MP4 frames. This mixed q8/BF16
+package stayed visually in the same family with mean frame MAE `1.66` versus source/BF16.
 
-Use the `mlxgen` command and Python import path for new MLX-Gen projects.
+`Storage` is the Hugging Face repository total. `Logical Model` is the loaded Wan transformer plus
+VAE tensor footprint measured from MLX arrays. `Full-Process Physical Peak` is Darwin
+`phys_footprint` sampled from model initialization through MP4 save and health validation.
+
+Validation assets:
+
+- [contact-sheet.png](validation/ti2v5b-clean/contact-sheet.png)
+- [metrics.json](validation/ti2v5b-clean/metrics.json)
 
 ## Usage
 
@@ -58,20 +84,23 @@ mlxgen download --model AbstractFramework/wan2.2-ti2v-5b-diffusers-8bit
 
 mlxgen generate \
   --model AbstractFramework/wan2.2-ti2v-5b-diffusers-8bit \
-  --task text-to-video \
-  --prompt "Your video prompt here" \
+  --prompt "A short cinematic video of a glowing orange glass sphere floating above calm teal water, soft reflections, gentle camera movement" \
+  --negative-prompt "" \
   --width 1280 \
   --height 704 \
-  --frames 121 \
-  --steps 50 \
+  --frames 17 \
+  --steps 20 \
   --guidance 5 \
   --fps 24 \
-  --seed 42 \
+  --seed 321 \
   --output video.mp4
 ```
 
+TI2V-5B also supports first-frame image-to-video in MLX-Gen when one input image is supplied.
+
 ## Attribution
 
-MLX-Gen is based on [mflux](https://github.com/filipstrand/mflux) by Filip Strand and the original mflux contributors. This model card is generated by MLX-Gen so derived checkpoints keep that attribution visible.
+MLX-Gen is based on [mflux](https://github.com/filipstrand/mflux) by Filip Strand and the original
+mflux contributors.
 
 Quantized and contributed by [@lpalbou](https://huggingface.co/lpalbou).