Datasetpublic

Limbicnation/dual-stream-image-prompts

Dual-Stream Image Prompts Multi-dialect image-prompt SFT dataset for training an LLM to route prompts to the right diffusion model at inference time. Given a concept and a target_model, the model learns to emit the correct prompt dialect (FLUX T5-XXL prose, SDXL dual-clip tokens, a compact caption, or steering modifiers). Routing lives in the instruction prefix, not in a nested output object — keeping the LoRA's task simple and maximizing structural diversity for generalization.… See the full description on the dataset page: https://huggingface.co/datasets/Limbicnation/dual-stream-image-prompts.

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes37downloads

Dataset Card

Dual-Stream Image Prompts

Multi-dialect image-prompt SFT dataset for training an LLM to route prompts to the right diffusion model at inference time. Given a concept and a target_model, the model learns to emit the correct prompt dialect (FLUX T5-XXL prose, SDXL dual-clip tokens, a compact caption, or steering modifiers).

Routing lives in the instruction prefix, not in a nested output object — keeping the LoRA's task simple and maximizing structural diversity for generalization.

Dataset at a glance

—22,878 rows (20,590 train / 2,288 validation, 90/10 split, seed 42)
—4 prompt dialects (target_model)
—Derived from `Limbicnation/images-diffusion-prompt-style-v2` (6,722 human-judge-validated prompts, mean judge score 4.875/5)
—Generated with gemini-2.5-flash-lite, rule-gated per dialect

Schema

Column	Description
`original_concept`	The source concept / style name
`target_model`	Dialect: `flux_t5` \	`sdxl_dual_clip` \	`compact_caption` \	`steering_modifiers`
`instruction`	Dialect-specific instruction prefix + concept
`response`	The prompt rendered in the target dialect
`text`	Full Qwen2.5 chat-template render (system + instruction + response) — train-ready

Dialects

`target_model`	Instruction prefix	Response shape	Rows
`flux_t5`	"Generate a FLUX (T5-XXL) image prompt for:"	dense natural-language prose, 40–120w	6,077
`sdxl_dual_clip`	"Generate an SDXL image prompt for:"	front-loaded comma tokens, 15–60w	6,322
`compact_caption`	"Write a compact descriptive caption for:"	25–50w enriched caption (≥1 lighting term)	5,728
`steering_modifiers`	"List image steering modifiers for:"	5–8 comma-separated tags	4,751

How it was built

A transform-assisted pipeline (scripts/generate_dual_stream_dataset.py in prompt-lora-trainer):

1.Each vetted v2 response is reused directly as the flux_t5 dialect (already FLUX-T5-shaped: median 81w, 90.6% within 40–120w).
2.One LLM call per concept derives the other three dialects (compress → SDXL, synthesize → caption, expand axis → modifiers).
3.Each dialect is gated independently (partial salvage — a failed dialect does not discard the others). ~15% per-row attrition; 22,878 of 26,816 candidates passed.

Quality gates (calibrated against the v2 source)

Dialect	Hard rules	Soft (warn-only)
`compact_caption`	25–50w · ≥1 lighting term · no filler buzzwords	camera term
`steering_modifiers`	5–8 tags · no duplicates · no generic terms	lowercase
`flux_t5`	40–120w · natural prose (no lists)	—
`sdxl_dual_clip`	15–60w	front-loaded subject

The camera-term check is intentionally soft: only 53.5% of the source's 4.8★ prompts contain one, so a hard gate would have discarded ~47% of high-quality content.

Usage

python

from datasets import load_dataset

ds = load_dataset("Limbicnation/dual-stream-image-prompts")

# Train on the chat-rendered `text` field directly (Qwen2.5 template)
print(ds["train"][0]["text"])

# Or filter to a single dialect
flux = ds["train"].filter(lambda r: r["target_model"] == "flux_t5")

At inference, prompt the fine-tuned model with the dialect prefix, e.g. "Generate an SDXL image prompt for: <your concept>".

Limitations

—Synthetic (Gemini-generated); dialect splits are uneven (steering_modifiers is the smallest at 4,751 rows due to stricter tag-count gating).
—flux_t5 rows are lightly-processed copies of the source dataset's prompts.
—English only. Concepts inherit the source dataset's style/subject distribution.

Source & license

Derived from Limbicnation/images-diffusion-prompt-style-v2. Released under Apache-2.0.