Multy_prompts
A ComfyUI node that turns one character box, one shared prefix, and one prompt wall into one finished prompt, one seed, and one plate name per image. The node does not load a model and does not sample. Your existing graph runs once per plate.
Search for Multy_prompts. It is under utilities → prompt.
Install
cd ComfyUI/custom_nodes
git clone https://github.com/JDJoe/ComfyUI-Multy_prompts.git
Restart ComfyUI. Search for Multy_prompts. Leave the folder name as ComfyUI-Multy_prompts.
The graph also uses three built-in nodes:
| Node | Menu |
|---|---|
| Start Loop | utilities → looping |
| End Loop | utilities → looping |
| Get Item From List | utilities |
What to add
Keep the model loader, CLIP, VAE, LoRAs, empty latent, negative prompt, sampler, decode, and Save Image that you already use. Add these four nodes:
- Multy_prompts
- Start Loop
- Get Item From List
- End Loop
Set Start Loop mode to List. On the sampler, set the seed control to fixed. randomize replaces the seed from this node.
Leave Start Loop parent_iteration and initial_iteration_value empty. Those are for a loop inside another loop, and for a value carried from pass to pass.
Connections
Multy_prompts has three outputs. Start Loop can walk only one list, so the prompts use the loop. The seeds and the plate names each use their own Get Item From List, both indexed by iteration_index.
| From | To |
|---|---|
Multy_prompts prompts |
Start Loop list |
Start Loop list_item |
positive text encoder prompt |
Multy_prompts seeds |
Get Item From List list |
Start Loop iteration_index |
Get Item From List index |
| Get Item From List output | sampler seed |
Multy_prompts plate_name |
a second Get Item From List list |
Start Loop iteration_index |
that Get Item From List index |
Save Image images |
End Loop termination |
Drag list_item onto the prompt text box, and drag the Get Item From List output onto the sampler seed box. That turns each widget into an input.
The positive encode, the sampler, the VAE decode, and Save Image sit between Start Loop and End Loop. One queue then writes one image per plate.
Model, CLIP, VAE, LoRAs, latent size, and the negative prompt stay outside the loop. They load once and are reused for every plate.
Multy_prompts.prompts ──► Start Loop.list
Start Loop.list_item ──► positive text encoder.prompt ──► sampler ──► VAE Decode ──► Save Image
│
Multy_prompts.seeds ──► Get Item From List.list │
Start Loop.iteration_index ──► Get Item From List.index │
Get Item From List ──► sampler.seed │
│
Multy_prompts.plate_name ──► Get Item From List.list │
Start Loop.iteration_index ──► Get Item From List.index │
Get Item From List ──► Concatenate Text.string_b ──► Save Image.filename_prefix │
▼
End Loop.termination
Prefix
Optional. Style text, LoRA triggers, or anything else that stays the same. It is pasted in front of every plate. A blank box adds nothing.
35mm photo, natural light, fine grain
CHAR tokens in the prefix are replaced the same way as in a plate. The plate id is still not included.
Characters
Optional. Leave the box empty and the prompts go out unchanged.
One character per block. A blank line separates blocks. The block starts with the token. The description can follow on that same line, or on the lines under it. Write a phrase, because that phrase is pasted in place of the token.
CHAR1 A woman in her thirties with short black hair and a grey coat.
CHAR2
A tall man with a scar on his left cheek.
Tokens are CHAR plus a positive integer: CHAR1, CHAR2, CHAR10. CHAR10 is not treated as CHAR1. A token with no character block is left as written. A prompt that never names a character is not rewritten.
Prompts
Same headers as a PlatePress shot wall. A plate id on its own line starts the next image. Blank lines inside a plate stay in the text. The id is not sent to the sampler. It is sent on plate_name, one name per image, in wall order. p04_ comes out as p04. p02_letter comes out as p02_letter.
To use that name as the file name, connect the Get Item From List item to Concatenate Text string_b. Type the folder in string_a. Set the delimiter to /. Convert Save Image filename_prefix to an input and connect the concatenated text there. Folder inkconf and plate p04 save as output/inkconf/p04_00001_.png. Save Image still adds the _00001_ counter. An empty folder with delimiter / becomes a leading slash, which Save Image rejects. Type the slash in the folder box (inkconf/) and leave the delimiter empty when that box may be blank. An empty box then saves p04_00001_.png.
p01 and p01_station both count. A leftover underscore, as in p04_, counts as p04. p1 and p001 are the same plate. Text after the id on that same line is the first line of the prompt.
p01
CHAR1 stands in the rain outside a station.
p02_letter
CHAR1 hands CHAR2 a letter in a dark hallway.
The hall is dark. A single bulb hangs over the door.
Seeds
One line per plate, in the same order as the plates. An integer keeps that seed. An empty line chooses a seed when you queue. An empty box chooses a seed for every plate. Fewer lines than plates keep the seeds you wrote and choose a seed for each plate after that.
657539441289641
1001
The blank second line is a chosen seed for p02. Two lines and eleven plates keep those two seeds and choose the other nine.
Sample workflows
Both samples use Start Loop in List mode, a fixed sampler seed, and 768×1376. There is no reference image. The LoRA stack is Lora Loader Stack (rgthree) from rgthree-comfy.
krea2_t2i_multy_prompts.json is the Krea 2 sample. The positive encoder is Text Encode Qwen Image Edit Plus. Its image slots stay empty. The plate from Start Loop connects to its prompt. The negative stays on CLIP Text Encode, taken straight from the CLIP loader. The model is KREA2/krea2_turbo_fp8_scaled.safetensors. Slot 1 of the LoRA stack is KREA2/Krea_2_Xenoflora_Trees_tcrzit_BF16-3K.safetensors at 0.75.
moss_t2i_multy_prompts.json is the MOSS-Image sample. The positive prompt is CLIP Text Encode, fed by Start Loop. Skip Layer Guidance sits between the LoRA stack and the sampler. plate_name goes through a second Get Item From List into Concatenate Text string_b, and that text is the Save Image filename. The folder in string_a is MOSS01_Multy_prompt01/, with an empty delimiter. Plate p01_awning saves as output/MOSS01_Multy_prompt01/p01_awning_00001_.png. The model is MOSS/moss_image_sft_int8_convrot.safetensors, the CLIP is MosTx/qwen3.5_9b_w4a8_gptq.safetensors, and the VAE is moss_vae/flux2-vae.safetensors. The stack slots are MOSSi/MOSSi-inkconf-cmx01-88.safetensors at 0.8 and MOSSi/MOSSi-inkconf-raw-01.safetensors at 0.75.
The identity-edit graph, which keeps a person from a reference photo, is krea2-ref-image-multy-prompts.
Not in this version
Reference images are not an input on this node. A later version can attach a character image to every plate that names that character, and a scene image to one plate. The text path above does not have to change for that.