Fix broken PhotoMaker v1 path: version-aware num_tokens + pass pm_version="v1" where the v1 checkpoint is loaded

[Linxiushen]

Problem

load_photomaker_adapter() defaults pm_version="v2" and then hardcodes self.num_tokens = 2. But the v1 flow the README still documents, both demo notebooks, and predict.py all download photomaker-v1.bin and call the loader without pm_version, so a v1 checkpoint is loaded down the v2 code path. Two failures follow:

1. Setup fails on the state dict (this is what issues RuntimeError: Error(s) in loading state_dict for PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken #194, The Notebook demo can't work #168, Error: photomaker_style_demo in qformer_perceiver.perceiver_resampler.layers.3.1.3.weigh #175 and error in running demo note book #193 report).
2. Even past that, the v1 ID encoder produces one embedding per ID image while the pipeline reserved two class-token slots, so PhotoMakerIDEncoder.forward raises:

RuntimeError: Sizes of tensors must match except in dimension 1.
Expected size 2 but got size 1 for tensor number 1 in the list.

So fixing only the version argument just moves the crash later; both parts are needed, which is why this is one PR rather than two.

Reachability, stated plainly: the README quick start (lines 108-160) and the two notebooks it links are the affected paths, and copying them fails immediately. I am not claiming the hosted Replicate endpoint is down: its image was built in 2024-01 from v1-era code and is frozen, and cog.yaml pins diffusers==0.25.0, which cannot even import today's photomaker/pipeline.py (diffusers.callbacks only exists from 0.28). The predict.py line is still worth fixing, but not because production is broken.

Fix

• self.num_tokens = 2 if pm_version == "v2" else 1, which restores the v1 semantics the code had before the v2 commit ([class_token] * num_id_images).
• Pass pm_version="v1" at the four call sites that load photomaker-v1.bin: the README snippet, both notebooks, predict.py.
• The same hardcoded num_tokens is fixed in pipeline_controlnet.py and pipeline_t2i_adapter.py for consistency. Stated plainly: those two have no v1 caller in the repo today (both inference_pmv2_* scripts fetch photomaker-v2.bin and use the v2 default), so they are latent, not currently broken.

The loader signature and its pm_version="v2" default are unchanged, so every v2 caller (the four inference_pmv2*.py scripts and gradio_demo/app_v2.py) is untouched.

Verification

zsh verify.sh in my scratch dir reproduces the whole thing; the important parts:

• v2 unchanged: 5 prompt templates x num_id_images in {1,2,3,4,5,8} (30 combinations), comparing clean_input_ids and class_tokens_mask between the pristine and patched trees: all identical. End-to-end forward through the real, unmodified PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken with a fixed seed: max absolute difference 0.0, torch.equal true.
• v1 fixed: with num_id_images of 1, 2, 4 and 8, the pristine tree raises the RuntimeError above every time; patched, all four return out=(1, 77, 2048).
• Notebook edits are surgical insertions into the original JSON (not a json.dump rewrite), so the diff is one added and one changed line each and both still parse. README.md is CRLF and the patch preserves that.
• ruff with the repo's pyproject.toml settings reports the same 221 findings before and after, with an identical per-rule breakdown; no new diagnostics.

No test file: the repo has none, requirements.txt has no pytest, and there is no CI, so testing this end to end would mean mocking the whole diffusers pipeline or downloading multi-GB weights. The differential script above covers it instead.

Note for whoever picks this up: open PR #200 touches photomaker/pipeline.py heavily, so if that lands first this needs a trivial textual rebase.

---

This fix was developed with AI assistance (Claude); the change was reviewed and verified locally before submission.