ComfyUI-RemoteTextEncoder

CLIPTextEncodeCoupleRemote

SDXL-style dual-encoder conditioning node. In SDXL, conditioning is produced by two CLIP models: • clip-l (ViT-L/14) – 77 tokens × 768-dim • clip-g (ViT-bigG) – 77 tokens × 1280-dim, also provides the pooled vector The combined conditioning tensor is the concatenation along the last axis, and the pooled output comes from clip-g only. Connect: RemoteDualCLIPLoader.clip → .clip (model 1 = clip-l, model 2 = clip-g)

Conditioning/Remote · 1 Eingabe · 2 Parameter · 1 Ausgabe

CLIPTextEncodeRemote

Drop-in replacement for CLIPTextEncode that sends the prompt to the Remote Text Encoder server and returns standard ComfyUI CONDITIONING. Connect: RemoteCLIPLoader.clip → CLIPTextEncodeRemote.clip primitive string → CLIPTextEncodeRemote.text CONDITIONING output → KSampler.positive / .negative

Conditioning/Remote · 1 Eingabe · 2 Parameter · 1 Ausgabe

LTXVRemoteCLIPLoader

LTX-Video 2.3 hybrid CLIP loader: Gemma-3-12B runs on the remote server, the ``text_embedding_projection`` linear layer loads from a local ``.safetensors`` file and runs on the ComfyUI machine. Why split? Gemma-3-12B is ~12 GB – offload it to a server with a big GPU. The projection weights (``ltx-2.3_text_projection_bf16.safetensors``) are ~1.4 GB; keeping them local avoids syncing large files to the server while still saving the dominant VRAM cost. Required projection .safetensors keys -------------------------------------- • ``text_embedding_projection.weight`` shape [3840, 188160] (= 3840 × 3840 × 49) Connect: LTXVRemoteCLIPLoader.clip → LTXVTextEncodeRemote.clip

Conditioning/Remote · 1 Ausgabe

LTXVTextEncodeRemote

LTX-Video 2.3 text conditioning node. LTX-Video 2.3 (Lightricks) uses two text encoders: • T5-XXL (primary cross-attention) – max 256 tokens, 4096-dim • Gemma-3-12B (secondary guidance) – max 512 tokens, variable-dim This node calls the server’s /comfy/encode/ltxv endpoint in a single HTTP request and assembles the ComfyUI CONDITIONING expected by LTX-V sampler nodes: [[t5_cond [1, T5, 4096], { "pooled_output": t5_pooled [1, 4096], "gemma_embeds": gemma_cond [1, Tg, D], "gemma_pooled": gemma_pool [1, D], }]] Connect: RemoteDualCLIPLoader.clip → .clip (model 1 = T5-XXL, model 2 = Gemma-3-12B) text prompt → .text CONDITIONING output → LTXVScheduler / LTXVSampler

Conditioning/Remote · 1 Eingabe · 3 Parameter · 1 Ausgabe

RemoteCLIPLoader

Replacement for the built-in CLIPLoader / DualCLIPLoader nodes. The ``model_name`` dropdown is populated live from the server’s GET /v1/models endpoint using the URL stored in ``rte_config.json``. Run the **Refresh Remote Models** utility node or press F5 in ComfyUI to update the list after loading a new model on the server. If a model you need is not in the dropdown yet, type its name into the ``custom_model`` field – this overrides the dropdown selection. Outputs a REMOTE_CLIP handle that stores the server connection parameters. No network call is made at load time.

Conditioning/Remote · 1 Ausgabe

RemoteDualCLIPLoader

Loads two text encoders from the same server and outputs a single CLIP object — exactly like ComfyUI's built-in DualCLIPLoader. Wire the output to: • CLIPTextEncodeRemote – encodes the same text with both models (SDXL default) • CLIPTextEncodeCoupleRemote – SDXL with separate text_l / text_g prompts • LTXVTextEncodeRemote – LTX-Video T5 + Gemma encoding • Any standard ComfyUI CLIPTextEncode node

Conditioning/Remote · 1 Ausgabe