Finetuning and training takes a lot of time and energy: If you enjoy what I do and want to help, why not donate? Click here!
(Updates and revisions to DREAD will always be free! This is just if you want to!)
A custom build model using SD1.5 as a base. Trained on 30k images ove it's 4 versions, (make that 35000 over 5) of horror and psychological terror, with a very graining and dark aesthetic. The images were hand sorted to make sure they fit the style and theme I was looking for, and after auto-tagging were then manually tagged with more clear identifiers and themes to help push generation even further.
If you are looking for crisp and clean images, go elsewhere. This is one grainy boi.
All sample images are without any post processing, upscaling, or cherry picking. Just random / same seeds as previous versions and posting.
V5 now done! (lora here - is large because when extracting at a lower DIM it wouldn't work well)
Dread-v5 adds more varied images and much more complex tagging for all of them, at a higher resolution. This will be the final Dread version, at least for SD1.5, since there is little else to be done to the base model. From here on out, any additions will be via "expansion pack" lora.
USE TIPS
V5 has many more keywords trained into it than the older versions, so use is both much more fine tuned but also a bit more finicky. The keywords below are ones that can help evoke horror feelings and moods, but are far from necessary:
found footage, disposable camera, trailcam, liminal, abandoned, hallway, house, forest, swamp, vines, roots, branches, dendrophobia, body horror, visceral, circulatory system, viscera, organs, blood, injury, laceration, gore, body horror, boil, rash, peeling flesh, fleshy, wet, water, drowning, burning, fire, flames, injury, plants, mushroom, fungus, diagram, illustration, tome, grimoire, arcane, sigil, magic symbol, experiment, medical, research, lab, malpractice, contained, restrained, chains, cables, cord, torture instrument, dungeon, mansion, manor, estate, swamp, entity, creature, being, monster, animal, glowing eyes, multiple eyes/limbs/mouths/etc, trypophobia, holes, torture, suffering, pain, despair, open mouth, sharp teeth, spikes, religious, demonic, ritual, cult, demonic, angelic, and many more...
Push the style further
Use DREAD in your prompt, all caps, to reinforce the horror aesthetic. Only the best images were tagged with this, and it can change your generations greatly. This is not a required tag, however! All images for training contained the word 'horror' in their tags.
One of the sample images is a grid showing a few prompts of the same seed, but using different samplers- just so you can see how each one influences the output and reading of this model.
Description
6.5k images trained on top of the initial set, more varied data and better contrast ranges. Model is harder to work with due to this, however.
FAQ
Comments (9)
Been using Dread V1; really looking forward to testing this. Extremely effective for img2img work, including a lot of non-horror subjects!
I'm getting only black outputs, with "A tensor with only NaNs was produced in the Unet" errors from A1111.
weird. did you download only the LoRA and are using it as a normal model?
@hobolyra Nope. In fact, I've tried both the lora and the regular v2 model. Also tried with --no-half.
@Baughn Turns out this is a bug currently with CivitAI downloads, and thus nothing I can do on my end about it. https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/6923#issuecomment-1489520816
They apparently fixed it a few hours ago, so redownload the model and hopefully that will help.
This model is great for a lot of things. I do wish it had more training on mechanical objects and details and some specific words (for example, "blister", "sporepod", "chitin"); I've gotten some really cool, gnarly results out of this, when doing img2img work, though!
I'll take that into consideration for future tweaking, thanks!
@hobolyra I'm using V4 now, and digging it. "chains", "gears", "hydraulics", various words for metallic stuff- "rusty" "distressed" "sharp" etc. might help. I'm getting really good results with this for things where I want grimdark / post-apocalyptic mechanical mixed with biomechanical looks, and I just want more, lol! Anyhow, this remains one of the best general-purpose models I've used, for stuff where I don't want photoreal and want subtle, washed-out hues. Great for img2img.
@coolmonkey Glad it's working for you! I need to find a nice balance between finetuning new / more refined concepts, and it overwriting old ones. Still figuring out how I'm going to approach it.
One thing I found can make an even greater style change is also using the extracted Lora on TOP of this same model. Sometimes it's way too much, other times it actually augments things very well.