CLIP Text Encode (Kandinsky 5)

Kandinsky 5 is the open image model from the Fusion Brain team (the Russian lab behind the Kandinsky line), and it takes its prompts through two encoders - an old friend and a new one. CLIPTextEncodeKandinsky5 has both boxes: clip_l and qwen25_7b. It landed in core in December 2025 with Kandinsky 5 support (the lite and pro variants).

The two boxes

• clip_l - the familiar CLIP-L encoder. Short, keyword-style content.
• qwen25_7b - a Qwen2.5-7B text encoder. This is the LLM doing the heavy lifting, and it's the reason Kandinsky 5 understands whole sentences instead of just tag lists.

The node tokenizes each box into its own slot - the Qwen text goes into tokens["qwen25_7b"] - and encodes them into a single CONDITIONING. The structure mirrors what the other LLM-equipped models do, just with a different cast of encoders.

What to actually type

The split is the practical part. clip_l wants the compact, descriptive core ("a red fox in snow"). The qwen25_7b box wants the full instruction - the scene, the style, the lighting, the "make it moody" part. Because it's an instruct-tuned LLM reading it, plain natural language beats tag soup, and long detailed prompts are exactly what this model was built to consume. Leaving the Qwen box empty is the classic way to get a Kandinsky 5 image that barely acknowledges your prompt.

The gotchas

The clip object must contain both encoders. Load Kandinsky 5's CLIP with the right loader or the node errors on the missing qwen25_7b slot - same story as every multi-encoder model, but worth repeating because Kandinsky is obscure enough that copy-pasted workflows with a generic CLIP loader get shared around. Second: if you're running Kandinsky 5's image-to-video side, there's a separate node for that; this one is purely the text conditioning.

Honest context

Kandinsky 5 isn't the model everyone's talking about, and that's a shame in some ways - the "forgotten model" angle is real. It's open, it's LLM-prompted, and its conditioning path is simple once you know it. But the ecosystem around it is thin, so expect to troubleshoot with fewer community answers than you'd get for Flux or SDXL. The node itself is dead simple: fill both boxes, wire the CONDITIONING into the KSampler, done. It ships with ComfyUI core - no install, and the only real setup is having the two encoder files in models/text_encoders (or models/clip) and loaded into the clip the checkpoint expects.