Load CLIP (Downloaded)

Load CLIP (Downloaded) is the piece of ComfyUI Model Batch Downloader that loads the text_encoders category - the file that turns your prompt into the conditioning your diffusion model understands. In split-model setups it's the second load, the companion to Load Diffusion Model (Downloaded). Its outputs are CLIP, and you wire that into your CLIP Text Encode nodes exactly like you would from a normal Load CLIP.

Why this node, not just Load CLIP

Because Load CLIP reads from a folder dropdown, and the file isn't in a folder yet - it's in a download_result from the batch downloader. More importantly, the type dropdown here isn't a cut-down list. The pack's README is explicit that this node uses the same type choices as ComfyUI's standard Load CLIP, and that set has grown with the ecosystem: alongside the old stable_diffusion and sd3 you'll find ltxv, wan, cosmos, lumina2, chroma, and the newer krea2, ideogram4, and flux2. The README's Krea 2 example is built entirely around this: download the diffusion model and VAE, download the text encoder into text_encoders, then select krea2 here.

That list matters because a modern text encoder isn't a generic CLIP anymore. Some families use a Qwen-3 or similar LLM as the encoder, and the whole model's behavior depends on the type matching what you downloaded. Get the family right or the conditioning is garbage - this is not a "any type works" situation.

Inputs and outputs

• download_result - from the download node.
• id - the manifest ID of the encoder file (defaults to the filename stem if unset).
• type - the model-family dropdown described above. This is the one you must set correctly.
• device (optional) - default or cpu. Forcing cpu is a legit move when the text encoder is the VRAM bottleneck: encoders are frequently the piece that decides whether anything fits on smaller cards, and they tolerate being offloaded better than the diffusion model. It'll be slower per encode, but it frees VRAM for the big file.

Output: CLIP, feeding both your positive and negative CLIP Text Encode nodes.

The workflow shape

Anima-style setups read like this: the sampler takes MODEL from the diffusion loader, both CLIP outputs (this node's) into the two text encodes, and VAE from the VAE loader into decode. If that's the workflow you're assembling, remember the downloader's category rule - the encoder must land in text_encoders or this loader will refuse it with a "requires text_encoders" error rather than loading the wrong thing. It's a strict guard, and it's your friend.

Install and troubleshooting

The shared pack install: ComfyUI Manager search for "ComfyUI Model Batch Downloader", or git clone https://github.com/watarika/ComfyUI-Model-Batch-Downloader.git into ComfyUI/custom_nodes, then restart. No pip dependencies; just confirm aria2c is on PATH for the downloads.

The failures you'll meet are the pack's usual trio: an id that doesn't exist in the manifest (double-check spelling or whether the download actually completed - interrupted downloads resume next run via the .aria2 sidecar), a category mismatch if you filed the file under the wrong model_type, and the genuinely confusing one - a type that doesn't match the encoder file, which usually surfaces not as an error but as garbage conditioning or odd generations. When the image looks semantically lost, check the type dropdown before the seed.