CLIP Text Encode (ablated Lumina2)

Read the node title and you'll think this is a CLIP text encoder, same as the one in every SDXL workflow. It isn't. Lumina Image 2.0 - the Apache-licensed DiT from Alpha-VLLM that refused to die after Flux took over - doesn't use CLIP at all. Its "clip" is a Gemma-2-2B instruction-tuned LLM, and that difference is the whole reason this node exists.

What it actually is

This is a near-drop-in replacement for ComfyUI's built-in CLIPTextEncodeLumina2 node, and I mean near literally: the author forked the core node's code and changed exactly one thing. The stock node gives you two system prompts, Superior and Alignment - the two Lumina2 ships with. This one adds a third, called Ablated, which is the standard "uncensored assistant" jailbreak framing, the kind that insists all refusal vectors have been removed from its programming.

That's it. That's the pack ("L2UC" = Lumina 2 Unchained). One node, one extra dropdown entry. The GitHub README is untouched cookiecutter scaffolding, so don't go looking for documentation there - the node's own tooltip and source tell you everything.

How it works

Because the encoder is a chat-tuned LLM, prompting it is chat, not tag soup. The node builds one string:

{system prompt} <Prompt Start> {user prompt}

then tokenizes and encodes it into a CONDITIONING tensor, exactly like the core Lumina2 node. The <Prompt Start> marker is Lumina2's prompt-format separator, not something this pack invented. The system prompt you pick is genuinely load-bearing here - on an instruct model, that preamble steers the whole encoding, which is why "just add more tags" people get sad on Lumina2 and why a third option can matter at all.

The inputs and output

Three inputs, and you'll set all of them:

• system_prompt - a dropdown: superior, alignment, or ablated. The author's tooltip explains all three in full. Start with alignment; switch to ablated only when the encoder is dodging content you want.
• user_prompt - your actual prompt, multiline and natural-language. Lumina2 likes sentences; quality tags like "masterpiece" do almost nothing here.
• clip - the CLIP-model input, fed from the loader. For Lumina2 you load the Gemma encoder, not an SDXL clip.

The single output is CONDITIONING, which wires straight into whatever sampler or CFG-guided node your workflow uses - same as any other text-encode output.

The honest take on "Ablated"

Set expectations before you get excited: this is not an ablated model. No weights were touched. It's a jailbreak system prompt - a piece of text telling an instruct LLM to believe it has no refusals. Sometimes that genuinely loosens the encoder's self-censorship; for the vast majority of benign prompts, Alignment will produce the same image. Treat it as "maybe get fewer refusals," not "new capabilities unlocked." And if the point is just that the word "CLIP" bugs you - you're right to be annoyed, but that's a ComfyUI-wide naming quirk, not this node's fault.

Installing it

L2UC ships no runtime dependencies of its own, but it needs a ComfyUI recent enough to have native Lumina2 support (added February 2025 - update ComfyUI first). Then:

cd ComfyUI/custom_nodes
git clone https://github.com/I-ShadowStar/L2UC

restart ComfyUI, and it appears as CLIP Text Encode (ablated Lumina2) under conditioning. Or skip the terminal and search "Lumina2Unchained" (or "L2UC") in ComfyUI Manager.

The node itself is free, but the model is the real download. You'll need the Lumina Image 2.0 model file, the separate gemma_2_2b_fp16.safetensors text encoder, and the 16-channel FLUX VAE - all from the Comfy-Org repackaged repo. The text encoder is a separate file and a separate download, and forgetting it is the classic "no valid clip" trap.

Common issues

• "clip input is invalid: None" - the node's own error message, and it means your checkpoint loader didn't produce a usable encoder. For Lumina2 you generally don't load a checkpoint; you load the diffusion model and text encoder separately.
• No Lumina2 clip type in the loader - your ComfyUI is too old. Update it.
• Weird gibberish images - you loaded the wrong encoder (e.g. an SDXL CLIP) into the clip slot. Lumina2 needs Gemma.

For a model that spent early 2025 being written off, Lumina2 has a stubborn niche - Apache license, open training, good prompt comprehension. This node makes the slightly awkward part (its chat-format encoder) a little more flexible, and it's honest about what it adds.