Publish adapter: higgsfield

README.md

@@ -0,0 +1,41 @@
+---
+license: mit
+base_model: stabilityai/stable-diffusion-3.5-medium
+library_name: diffusers
+tags:
+  - sd3.5
+  - adapter
+  - higgsfield
+inference: false
+---
+
+# phenomenalai/sd3-token-mod-1024
+
+Adapter for `stabilityai/stable-diffusion-3.5-medium` trained with Higgsfield.
+
+## Usage
+
+```python
+import torch
+from diffusers import StableDiffusion3Pipeline
+from higgsfield.adapters.token_mod import GlobalTokenModulator
+from huggingface_hub import hf_hub_download
+
+base_model = "stabilityai/stable-diffusion-3.5-medium"
+repo_id = "phenomenalai/sd3-token-mod-1024"
+
+pipe = StableDiffusion3Pipeline.from_pretrained(base_model, torch_dtype=torch.bfloat16).to("cuda")
+# download adapter
+adapter_path = hf_hub_download(repo_id, filename="token_mod.pt")
+# load adapter
+mod = GlobalTokenModulator(num_tokens=333, embed_dim=4096, out_channels=pipe.transformer.config.in_channels)
+mod.load_state_dict(torch.load(adapter_path, map_location="cpu"))
+mod.to("cuda").eval()
+# Now use your own inference loop adding bias from `mod` like in your codebase.
+```
+
+## Files
+- `token_mod.pt` adapter weights
+- `run.json` metadata
+- `manifest.md` human summary
+

manifest.md

@@ -0,0 +1,14 @@
+# Higgsfield Run Manifest
+
+Model: stabilityai/stable-diffusion-3.5-medium
+Epochs: 2
+Batch size: 16
+Precision: bf16
+
+## Artifacts
+- Final adapter: runs/sd35_token_mod/final/token_mod.pt
+- Run JSON: runs/sd35_token_mod/final/run.json
+
+## Notes
+- Trained adapter: GlobalTokenModulator (token-to-latent bias)
+- Base components (VAE/Transformer/Encoders) were frozen.

run.json

@@ -0,0 +1,6 @@
+{
+  "model_id": "stabilityai/stable-diffusion-3.5-medium",
+  "epochs": 2,
+  "batch_size": 16,
+  "precision": "bf16"
+}

token_mod.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4107a373703cb581741aa06724fc6ffa3bd3194e1618ffa10e2b771233bdb802
+size 5753711