fix: handle missing Flux VAE config values and None prompt inputs

[octaviomejiadiaz]

Problem

When Flux is loaded from a ComfyUI single-file checkpoint (.safetensors), the bundled VAE often omits scaling_factor and shift_factor from its config. This causes a crash during sample generation:

`
TypeError: unsupported operand type(s) for +: 'Tensor' and 'NoneType'

diffusers/pipelines/flux/pipeline_flux.py line 1010

latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
`

A second related issue: the CLIP/T5 tokenisers raise a TypeError: TextEncodeInput must be Union[TextInputSequence, ...] when they receive None or False as the unconditional (negative) prompt, which can happen through the dropout path or when callers pass bare strings/None to encode_prompt.

Changes

oolkit/stable_diffusion_model.py

• After loading the VAE, inject the correct Flux defaults (scaling_factor=0.3611, shift_factor=0.1159) only when the values are absent - existing configs are untouched.
• Wrap the prompt argument in a list inside encode_prompt for the Flux branch, so a bare string or None never reaches encode_prompts_flux as a non-list.

oolkit/train_tools.py

• Normalise every element of prompts in encode_prompts_flux to a str, replacing None / False with '', before the tokeniser call.

Testing

Verified end-to-end with lux1-dev.safetensors loaded from a local ComfyUI checkpoint directory (24 GB VRAM, 8-bit quantisation, �damw8bit optimiser). Training now proceeds past the baseline sample generation step.

Notes

The correct Flux VAE constants (�.3611 / �.1159) match the values published in the official �lack-forest-labs/FLUX.1-dev model card.