.safetensors inputs from compatible architectures. Output is a new .safetensors file; a completed merge still needs loading and image-quality checks.Historical merge screen
Saient Merge, v1.0.4
This older screenshot predates the corrected diffusion-checkpoint picker. It is not a screenshot of dev.6.
Diffusion checkpoints, with tensor-only loading
Use .safetensors where possible. The worker also accepts tensor-only .ckpt/.pt dictionaries through PyTorch's weights_only=True loader. If safe loading rejects a file, the job fails visibly; there is no fallback to unrestricted pickle loading. Do not rename a file to bypass a format error.
The v1.0.9 report of unusable GGUF selectors remains a historical finding. It does not describe the corrected dev.6 picker.
What the merge worker implements
| Method | Formula | Use |
|---|---|---|
| Weighted sum | A × (1-w) + B × w | Interpolates matching tensors between two structurally compatible checkpoints. |
| Add difference | A + (B-C) × w | Applies a weighted difference between B and C onto A; model C is required. |
When a shared tensor has a different shape, the worker keeps A's tensor. Keys found only in A or only in B are copied into the output. That behavior prevents a shape exception, but it does not prove the resulting hybrid model is semantically sound.
Choose checkpoints that belong together
- Use the same architecture and closely related tensor layout—such as two checkpoints derived from the same SDXL base.
- Compare tensor-key and shape compatibility before treating a completed file as a valid model.
- Never merge a LoRA adapter as if it were a full checkpoint, and never merge GGUF language models with this diffusion worker.
- Keep the source files. A merge output is a new experiment, not a replacement for A, B or C.
- Review every source model's licence; blending weights does not remove upstream restrictions.
RAM and storage requirements
- GPU: not required by this tensor merge; it runs on CPU.
- Safetensors inputs: opened lazily, so input tensors are fetched as needed. The result still accumulates before save, making peak RAM roughly output-model scale plus working tensors.
- Tensor-only CKPT/PT inputs: loaded fully into RAM if safe loading accepts them; prefer safetensors for lazy input access.
- Disk: keep A, B, optional C and the new output simultaneously. Use a filesystem with enough free space for a failed/retried save as well.
Checked process behavior, limited model-quality evidence
Earlier native Linux development tests merged two tiny CPU tensor fixtures and independently checked the output values. They also exercised a failed save, cancellation without output, and a successful retry. These are historical integration checks, not a fresh dev.6 full-checkpoint benchmark.
The beta retains corrected model selection, visible worker errors and owned cancellation. A delayed cancelled request cannot start later or stop a replacement job. Output is saved through a temporary file before replacement.
What to do next
- Read the dev.6 Linux beta manifest and requirements before installing.
- Run compatible SDXL checkpoints locally.
- Train a LoRA locally when an adapter—not a full blended checkpoint—is the right tool.
- Read Saient setup help.
Keep source checkpoints and test the result.
The Linux beta corrects the picker and process handling. Keep A, B and C, use a distinct output path, and evaluate the merged model before depending on it.