stitchvm / README.md
HJGO's picture
Link published inference and training code
8087c08 verified
|
Raw History Blame Contribute Delete
1.61 kB
# StitchVM checkpoints
Pretrained StitchVM value-model checkpoints. See the [code and model tables](https://github.com/prs-eth/StitchVM) for the original twelve SD3.5 / FLUX combinations.
## HPSv3 and UnifiedReward-2.0 Qwen3.5
Epoch-5 EMA checkpoints with SD3.5 Medium. Each folder contains the trained
adapter and vision suffix, metadata, license notices, and SHA-256 checksums.
The upstream diffusion and reward models are required separately.
| Reward model | DiT → Encoder | Checkpoint | Details |
|---|:---:|---|---|
| HPSv3 | 5 → 2 | [Download](hpsv3_sd3.5_medium/model_trainables_ema.pt) | [Model card](hpsv3_sd3.5_medium/README.md) |
| UnifiedReward-2.0 Qwen3.5 27B | 3 → 2 | [Download](unifiedreward_qwen35_27b_sd3.5_medium/model_trainables_ema.pt) | [Model card](unifiedreward_qwen35_27b_sd3.5_medium/README.md) |
| UnifiedReward-2.0 Qwen3.5 4B | 4 → 2 | [Download](unifiedreward_qwen35_4b_sd3.5_medium/model_trainables_ema.pt) | [Model card](unifiedreward_qwen35_4b_sd3.5_medium/README.md) |
| UnifiedReward-2.0 Qwen3.5 9B | 3 → 2 | [Download](unifiedreward_qwen35_9b_sd3.5_medium/model_trainables_ema.pt) | [Model card](unifiedreward_qwen35_9b_sd3.5_medium/README.md) |
Training, inference, and evaluation code is available in
[StitchVM](https://github.com/prs-eth/StitchVM/blob/c7ad92a84ab11a56c5c380a823ea0357f71aa179/README.md). Start with the
[shared inference examples](https://github.com/prs-eth/StitchVM/blob/c7ad92a84ab11a56c5c380a823ea0357f71aa179/examples/README.md), which support all 16
released model combinations and download the matching EMA checkpoint automatically.