AI4ArtsEd T5-CLIP Fusion

AI4ArtsEd T5-CLIP Fusion fuses conditioning from a CLIP encoder with conditioning from a T5 encoder - the architecture behind Stable Diffusion 3, where CLIP-L handles the tag-ish short prompt and T5 handles the long natural-language one. The idea: interpolate the first 77 tokens of both embeddings (that's CLIP's whole world), then append T5's remaining tokens, which carry the long-prompt context CLIP never got to see. Alpha controls how much of that first 77-token block comes from CLIP (α=0) versus T5 (α=1).

Here's the honest part, straight from the source file's header comment: this node does not work as intended in this version. The author left a detailed tuning guide that reads like a field report from an experiment that half-escaped its cage:

• Negative alpha → increasingly "boring"; past -18.5 it "loses it completely"
• Around -1.5, T5 has no influence left
• -1.5 to +2 is the reasonable mixing zone
• Past +2 you start losing the prompt; past +7 it's gone; past +76 blackout

So you're not getting a clean CLIP/T5 blend - you're getting a strange, prompt-adjacent space that's fun to explore. The author says as much: "the output is fun to play with." If you treat it as an experimental node that occasionally produces fascinating accidents, you'll enjoy it. If you need reliable conditioning, reach for the standard encoder stack instead.

The inputs

• clip_conditioning - CONDITIONING from a CLIP encode.
• t5_conditioning - CONDITIONING from a T5 encode (Flux-style long prompt).
• alpha - default 0.75, range -75 to 75, step 0.005. That range is the tell: it's designed as a dial, not a ratio.

Output: one CONDITIONING, ready for the sampler.

Setting it up (for SD3-style models)

The node is built around the SD3 text-encoder trio. You need the three encoder files on hand:

clip_g.safetensors
clip_l.safetensors
t5xxlenconly.safetensors

Encode one prompt with CLIP (the clip_l path), one with T5 (the t5xxl path), and wire both into this node. Both embeddings must share the same dimension - the node checks and errors otherwise.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/joeriben/ai4artsed_comfyui_nodes
# restart ComfyUI

Or search "AI4ArtsEd" in ComfyUI Manager. Pure torch, no extra deps.

Common issues

The blend is gibberish outside -1.5..+2. That's expected per the tuning notes, not a malfunction. If you want a coherent image, stay in the reasonable zone; the outer ranges are for experiments.

Dimension mismatch. CLIP and T5 embeddings of different sizes will error out - they must match. The node returns your CLIP conditioning unchanged on any exception, so a silent "nothing happened" is your cue to check the console for the error print.

"Doesn't work as intended." Read that as a feature-flag warning, not a bug report. It's an academic, experimental node from an arts-education research pack, and it tells you so in its own source. Know what you're getting into and it's a fun sandbox; expect production conditioning and you'll be disappointed.