The LoRA field in job configs uses a combo string format:
"name:model_str:clip_str + name2:model_str:clip_str"
Parts can also be folder references like "FolderName/:0.90:0.90"
or random selection syntax "Folder/[3,0.5]".
_extract_required_models() was passing the entire combo string as a
single filename, causing WinError 123 (invalid path) because colons
are illegal in Windows filenames.
Now splits on " + ", extracts the filename before ":", and handles:
- Individual LoRA files: synced from master
- Folder references (trailing /): expanded locally via folder_paths
- Random syntax ([count,strength]): expanded locally via folder_paths
Same fix applied to distribution_manager._get_upcoming_models() which
builds the pre-fetch list for workers.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Move torch and safetensors imports inside huggingface_vae_decode() so
CivitAI and Distribution functions still work if those packages are
missing. Update stale "Uses urllib" comments to reference the gateway.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
All 6 urllib call patterns (download, register, claim, submit,
heartbeat, fail) now use gateway functions from network_utils.py.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Seed: expanded per-gen behavior + new full run seed behavior (random/increment/decrement before/after entire run)
- Canvas: fix flicker during simultaneous pan+zoom via requestAnimationFrame batching
- Upscaling: 3 session-level modes (HiRes fix, model upscale, model+HiRes) with tiled VAE support
- Cooldown: configurable GPU breaks between generations with optional VRAM clearing
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The worker's generate_image() call was only passing basic parameters
(seed, steps, cfg, sampler, scheduler, denoise, attention_mode) but
missing all advanced/model-specific parameters that the local path
passes. This caused workers to produce different images than local runs.
Added missing parameters:
- model_sampling_override (noise schedule for Flux/SD3/AuraFlow)
- model_sampling_shift (shift values for model sampling)
- model_sampling_flux_max_shift / model_sampling_flux_base_shift
- use_advanced_sampling (SamplerCustomAdvanced vs standard KSampler)
- advanced_guider / advanced_scheduler
- flux_guidance_value
- width / height (for Flux2Scheduler computation)
All parameters were already present in the job config dict from
expand_configs() but were never forwarded to generate_image().
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Worker now clones latent and frees UNet tensors + CUDA cache before
VAE decode, giving Dynamic VRAM clean memory for VAE loading.
decode_latent_with_vae now detects NaN in input/output, logs diagnostic
info, retries with float32, and handles bfloat16->numpy conversion.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Worker threads run outside ComfyUI's execution engine which wraps
everything in torch.inference_mode(). Without it, Dynamic VRAM's lazy
weight loading produces NaN during VAE decode, causing black images.
Matches ComfyUI's execution.py:703 pattern.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Worker downloads missing models from master via HTTP before processing.
Background pre-fetch thread downloads upcoming models during job execution.
Distribution manager scans pending queue for unique model references.
Job claim responses include upcoming_models list when sync enabled.
Master passes sync_models_to_workers flag to workers via start_worker.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
1. Worker registration now retries up to 5 times with 3s delay on 503,
handling the case where master is still starting up (e.g. pre-encoding).
2. Worker claim_job on 503 no longer immediately stops - gives master up to
5 poll cycles to become active before giving up (only when no jobs processed yet).
3. When master pre-encoding is enabled, worker skips build_prompt_with_triggers()
entirely (avoids CivitAI lookups and LoRA file hashing). Master now includes
the final prompt text (with triggers) in the job dict for metadata use.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
