pkmodel-flux-lora

A rank-16 LoRA for FLUX.1-dev that generates PowerPoint-style slide visuals: charts, diagrams, mind maps, podiums, timelines and similar template-like layouts with flat or gradient backgrounds.

It is the visual-style half of a slide-generation pipeline:

User prompt
   ├─► layout model (LLM) ──► layout JSON ──► python-pptx ──► editable .pptx
   └─► FLUX.1-dev + this LoRA ──► slide visuals / backgrounds / style reference

Trigger word

Start prompts with pkslide.

Usage

import torch
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16
).to("cuda")
pipe.load_lora_weights("mindfossil/pkmodel-flux-lora")

image = pipe(
    "pkslide SWOT analysis four quadrants, white background",
    width=1280, height=720,
    num_inference_steps=24,
    guidance_scale=3.5,
    generator=torch.Generator("cuda").manual_seed(42),
).images[0]
image.save("slide.png")

If the style is too strong or too weak, adjust the LoRA scale (for example pipe.set_adapters(["default_0"], adapter_weights=[0.7])).

Recommended settings: 1280x720, 24-28 steps, guidance 3.5.

Hosted inference

The Hub may offer to run this LoRA on third-party inference providers. Always include the pkslide trigger word in the prompt (the default example prompt does not), and keep the FLUX.1-dev non-commercial license in mind when using hosted services.

Prompt tips

Captions during training followed this pattern, so prompts in the same shape work best:

pkslide [subject], [categories], [tags], [colours], PowerPoint slide template professional presentation

Examples:

• pkslide bar chart slide dark navy background three series
• pkslide timeline process diagram white background five steps
• pkslide winners podium gold silver bronze trophies
• pkslide mind map blue silhouette head diagram
• pkslide 3D bar chart blue gradient ascending

Training details

Base model	FLUX.1-dev (quantized during training)
Framework	ai-toolkit by ostris
Method	LoRA, rank 16, alpha 16 (transformer only; text encoders frozen)
Steps	10,000
Batch size	1, gradient accumulation 4 (effective 4)
Learning rate	1e-4, adamw8bit
Noise scheduler	flowmatch
Precision	bf16 training, fp16 weights saved
EMA	enabled, decay 0.99
Resolution buckets	512, 768, 1024
Caption dropout	5%
Hardware	1x A100 40GB (Colab), roughly 5.5 hours

Dataset

• 11,496 slide images, each with a caption built from its metadata (title, categories, tags, colours).
• Images were padded to 1024x1024 using their own edge colour, so dark slides stay dark rather than gaining white bars.
• Trigger word pkslide prepended to every caption.

Intended use

• Generating slide-style visuals, backgrounds and style references for presentation tooling.
• Research and experimentation on template and diagram generation.

Limitations

• Output is a flattened image, not an editable slide. For real, editable decks use the layout model and a renderer (python-pptx) and treat FLUX output as imagery.
• Text inside generated images is often unreliable (garbled or misspelled), as with most diffusion models. Render text and chart data separately.
• Charts and diagrams are stylistic imitations; the numbers, labels and structure are not accurate data.
• Trained on a single stylistic family of slides, so behaviour outside that look (photographic scenes, non-slide content) is not improved and may be degraded at high LoRA strength.
• Because training images were padded to square, edge-colour bands can occasionally appear in non-square outputs.

License and data

• This LoRA is a derivative of FLUX.1-dev and inherits the FLUX.1-dev Non-Commercial License. Commercial use requires a separate licence from Black Forest Labs.
• The training images come from a commercial slide template library. The rights holder's terms may restrict redistribution or commercial use of models trained on that content. Review them before any public or commercial release.

Contact

Built by PK as part of AutonomyWorks.ai.