Start with the problem in front of you.

This historical status note records the author's April 14, 2025 README statement. It is not today's release announcement.

A wire reaching CLIPTextEncode does not prove that a valid encoder was loaded. Check upstream loading and whether the model uses a complete checkpoint or separate components.

Reference-image workflows need a vision encoder with the expected architecture. Distinguish it from the object feeding CLIPTextEncode.

Diagnose ControlNet compatibility, control-image preparation and conditioning separately; node-link type errors need their own check.

Identify the expected edge, pose or depth representation and trace how its conditioning reaches the sampler. Different control models are not one universal node.

A model name can cover bundled checkpoints, split weights, quantized variants, text encoders and VAEs. Preserve enough identity information to know which artifact a workflow actually used.

This reference covers native FLUX.1 text-to-image resource organization. dev, schnell, split weights and packaged FP8 checkpoints are different choices, not drop-in filenames for every graph.

GGUF is a container format, not a guarantee that an arbitrary chat-model file can serve as a diffusion model or text encoder in a particular ComfyUI node.

Identify the ordinary eight-step SDXL variant separately from CFG, SD1.5, SD3 and FLUX variants, with a file-level provenance record.

Identify the format and source of JSON, images or videos before installing nodes and weights. Static inspection helps planning but cannot replace execution in the target environment.

A downloaded IPAdapter file can exist on disk without satisfying a unified loader's preset matching. Identify the loader before renaming or relocating files.

IPAdapterUnifiedLoader is an entry point for understanding resource pairing, not an independent generator. Use the running node's definition for its precise interface.

An unexpected channel count or latent dimension calls for checking the model, VAE and latent-construction nodes together. Image and video latents do not all share the same structure.

Start with one compatible LoRA and a controlled comparison. A populated selector does not establish that the sampling path uses the modified weights.

A selectable LoRA can still target an incompatible model, or the workflow can bypass its modified outputs. Little visible change is an observation, not proof of a damaged file.

Establish the runnable range of a particular setup, not a universal minimum-VRAM claim derived from a model name.

This reference covers the open-weight native ComfyUI route, not remote Partner/API execution. Local operation depends on component pairing, core version and actual hardware.

Supported extra_model_paths configuration can reference existing model storage. It does not share Python environments, node registration or workflow versions.

A saved ckpt_name, lora_name or another selection is absent from the current allowed list. Check the input name first: this validation error can also concern non-file options such as samplers.

Start with native workflows and keep 2.1 separate from earlier Qwen checkpoints and hosted services. Find files, editing, transparency and licensing here.

Separate file loading, conditioning, sampling and decoding before changing the environment.

Select the diffusion model, conditioning encoder and VAE separately. Keep optional prompt enhancers and file-size estimates out of minimum-VRAM claims.

A trainer implementation is not a catalog of trained adapters. Track exact bases, adapter files and loader evidence separately.

Keep text-to-image and editing prompt rewriters separate from the base conditioning encoder.