═══════════════════════════════════════════════════════════════════════════════ FURRY ENHANCER VIDEO ═══════════════════════════════════════════════════════════════════════════════ ─────────────────────────────────────────────────────────────────────────────── IDENTITY ─────────────────────────────────────────────────────────────────────────────── Base model MiniMax H3 Generation mode Image-to-Video (I2VA, first frame), Text-to-Video (T2VA) Checkpoint fl2va (drives T2VA / I2VA / L2VA / FL2VA) Category Detail / anatomy fixer Author freek22 Source https://civitai.com/models/1782485 Provenance CivitAI · model 1782485 · version 3234825 ─────────────────────────────────────────────────────────────────────────────── WEIGHTS ─────────────────────────────────────────────────────────────────────────────── File Furry enhancer Video H3 V2.54.safetensors Version MiniMax-H3 V2.54 Size 887.3 MB Format SafeTensor / unknown precision SHA256 fc1e06352b7cc8689daa539ca3064af57d3265a4ada995c055f4fe66cdba3ab0 Published 2026-08-16 ─────────────────────────────────────────────────────────────────────────────── TRIGGER WORDS ─────────────────────────────────────────────────────────────────────────────── No trigger word. The LoRA applies as soon as it is loaded — strength is the only way to dial the effect up or down. ─────────────────────────────────────────────────────────────────────────────── RECOMMENDED SETTINGS ─────────────────────────────────────────────────────────────────────────────── Taken verbatim from the author's description: LoRA strength 1.0 Sampler steps 10 Resolution 480P Sampler euler ─────────────────────────────────────────────────────────────────────────────── ADOPTION & POPULARITY ─────────────────────────────────────────────────────────────────────────────── Downloads 19,301 Rating 100% positive (902/904) Comments 0 Published 19 days ago Download rate 60.1/day since release Adoption rank #4 of 114 in this collection (Top tier) Momentum rank #43 of 114 by download rate (Steady · fresh release) Adoption tier is the percentile of total downloads within this collection; momentum is the percentile of downloads-per-day since release. Momentum is ranked rather than measured against a fixed rate, because this base model is itself new and every LoRA looks fast on an absolute scale. Both numbers are CivitAI's, read on the scrape date at the foot of this file. ─────────────────────────────────────────────────────────────────────────────── CHAINING / STACKING ─────────────────────────────────────────────────────────────────────────────── Stacks with others yes What the author says (fact-checked against their own text): For the MiniMax-H3 version the author states only that it works with the turbo LoRAs, and that he uses mostly 10 steps in that setup ("i use a weigh between 0.5 and 0.9max! it works with the turbo loras i use mostly 10 steps"). No specific turbo LoRA is named. The two "Its also compatible with Ratatoskr" lines on the same model page belong to the Hunyuan and WAN 2.1 (I2V) sections of the description, not to the MiniMax-H3 LoRA; Ratatoskr is also the author's own checkpoint model rather than a companion LoRA, so it is base-model compatibility in any case. No combination to avoid, no required companion LoRA, and no load order are stated anywhere for this version. General MiniMax H3 rules that apply here (see REFERENCE.txt): • Load the LoRA with LoraLoaderModelOnly between UNETLoader and your sampler; stacking means chaining several of those nodes in series. • Stacks normally on top of a single acceleration/turbo LoRA; keep only one distill LoRA active in the chain. • Chain this only onto an fl2va checkpoint (T2VA / I2VA / L2VA / FL2VA). Loading it against ref2va can degrade output silently rather than erroring. ─────────────────────────────────────────────────────────────────────────────── PROS ─────────────────────────────────────────────────────────────────────────────── + Widely adopted — 19,301 downloads. [source: CivitAI stats] + Positively received — 902 thumbs-up, 2 down. [source: CivitAI stats] + Author documents use across several modes (I2V, T2V). [source: description] + Author published concrete recommended settings. [source: description] ─────────────────────────────────────────────────────────────────────────────── CONS & CAVEATS ─────────────────────────────────────────────────────────────────────────────── - Author labels this experimental — results vary shot to shot. [source: description] - Author references NSFW output — review before using in shared work. [source: description] - No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata] - Large at 887 MB — slower to load and heavier to stack. [source: file size] ─────────────────────────────────────────────────────────────────────────────── COMPATIBILITY WARNING ─────────────────────────────────────────────────────────────────────────────── MiniMax H3 ships two transformers that share tensor key names: the fl2va checkpoint (driving T2VA, I2VA, L2VA and FL2VA) and the separate ref2va checkpoint (driving Ref2VA). Loading a LoRA against the wrong family can silently degrade output instead of raising an error, so confirm the family above matches the checkpoint you have loaded. Likewise, LoRAs built for full checkpoints often throw adaln_proj shape errors on pruned ones — match the build to your checkpoint. ─────────────────────────────────────────────────────────────────────────────── AUTHOR'S DESCRIPTION ─────────────────────────────────────────────────────────────────────────────── This Lora is made to enhance the ability / quality to create anthro and furry but it works also for all kind of other fantasy creatures and animals too the range that it can handle for example the Wan 2.2 V2 Lora is from comic and drawing towards real photos . i made the lora for Hunyuan Video , WAN 2.1 and 2.2 (SFW and NSFW) Furry Enhancer Hunyuan V1 The Hunyuan Video version is trained for Text to Video (T2V) but i also tested it on Image to Video its also working with it very good but not made for that. All example images for the Hunyuan Video are made T2V Its also compatible with Ratatoskr Hunyuan Wan 2.1 I2v and T2V V1 The Wan 2.1 has 2 different models I2V and T2V the image to video is in most cases the model of choice for Wan for the Text to Video Variant please read the description. Its also compatible Ratatoskr WAN 2.1 (ITV) and will for sure increase the quality especially in the NSFW Area. I also tested the I2V Lora with Wan 2.2 it seems to enhance the result too but very experimental Wan 2.2 I2V and T2V V1 This enhancer version is for Wan 2.2 its splitted in basically 2 Loras: low and high noise both loras for I2V are trained on the same dataset. I its my own dataset before i trained with the wan 2.1 Lora but add more pictures and videos it's completely synthetic so no real persons or art is used for the dataset. whats the pros of this lora compared to the wan 2.1? - more consistent - more specimen better in sfw and nsfw i recommend using both loras and use the workflow in the training folder Note: your resolution needed to set between 0.4 (480P) and 0.92 (720p) overwise the lora generate backscreens or other failures. for strength i prefer the full 1.0 strength its tested its tested with the original wan 2.2 model also in a short test with my Ratatoskr WAN 2.2 Hybrid v1.0 WIP model (use only the high noise version with that model) for technical data about Wan 2.2 visit there official site here: V2 Furry Enhancer High Noise Changes compared to V1 - more consistent - more specimen - better overall movements - more stable overall - way better with comic and anime than V1 better in sfw and nsfw i recommend using the V1 Version of the WAN 2.2 I2V V1.0Low Noise Note: I used for example the old workflow for this model to compare V1 with V1 if you like the newest workflow download the training folder. Important for this workflow is to keep the resolution between 0.4 and 0.92 (480p - 720p) V3 High Noise and Low Noise This enhancer version is for Wan 2.2 its trained my own dataset it's completely synthetic so no real persons or art is used for the dataset. the biggest change in this dataset is its now completely trained on videos Changes compared to V2 - trained now on over 650 Videos - better overall movements - more emotions - way better with comic and anime than V1 and V2 better in sfw and nsfw for this Version there is also an newtrained Low Noise Model LTX2.3 Version (WIP Preview V0.1) this is an preview version of the first attempt to train the ltx 2.3 model with a very limited dataset so it's more like an technical demonstration / trial than a fully trained lora. So don't expect best results yet it's highly experimental. for workflow i have also added an experimental workflow. LTX2.3 Version V1 Lora Pack This is the First full Release Version of the LTX Version of Furry Enhancer Video. Better to say it's an model pack because V1.1 is made based on ltx 2.0 training setup so best for the LTX 2.0 model V1.215 its the main model trained specifically for the ltx 2.3 Version so start usually with that lora V1.22 More advanced in training but has sound issues. V1.3 experimental (wip) model trained in different training modes (heavy sound issues) you find the used workflows and the different model versions in the "training folder" the loras are tested on the dev version and the distilled variant. what has changed to preview? - Trained on the full video database (V1.2 and up) at around 700 videos instead of only 40 videos - Also trained on my photo database of ratatoskr overall over 10k pictures - Trained for i2v and t2v - better overall knowledge and movement . Note: this is still in development because most of the trained videos are based on my wan2.2 database so its planned to train an v2 this year with video output of that model. also i don't have my own workflow done yet so it depends on other workflow don by afroman4peace and akinson LTX2.3 Version V2.4 I have decided to retrain the dataset on a far advanced system now its trained on a server with an rtx6000 pro blackwell for more than a week 24/7 (and yes this was expensive) so what has changed from V1 to V2.4? - the trained video resolution is now at 720P constant - image training did improved from 1024p to 1536p - added 120 new sound videos for training especially for speech and jaw movements. - seperated videos with sound and not sound and trained in different databases - trained for T2V and I2V in stages so its nearly equally trained for both now - added new pictures to database too but more important overworked the captions fitting better for video training For settings and workflow i have added my example videos in the training folder the images and some videos has the metadata just put them into comfy on my workflows was the loral at best between 0.8 and 1.0 the lora is trained on the dev model but also tested for the distilled version for now still don't have my own workflow done yet so it depends on other workflow don by afroman4peace and akinson please support my work with an like and look on my other loras and models too the newest workflow download the training folder. in all videos the metadata is also included just download the video and put it into Comfy Important for this workflow is to keep the resolution between 0.4 and 0.92 (480p - 720p) if you like to support me please give a like and Check out my other models. Making such loras take a huge amount of cost and time please support my work with a like. and please check out my other models and loras too ;-) made by freek22 ─────────────────────────────────────────────────────────────────────────────── VERSION NOTES ─────────────────────────────────────────────────────────────────────────────── This is the second Version of Furry Enhancer Video for MiniMax-H3 all credits go to them first for such an amazing Video Model please read also there model license too So now to the lora this model is trained on around 450 videos and 7k images from my furry enhancer and ratatoskr database it's now trained on 96 rank on ab B200 and yes this is very expensive to do ;-) The Lora are trained with long videos up to 40 seconds and resolutions till 1024p but to V1.5 the videos where categorized in " length buckets" so very short videos are trained on 1024p and above where the really long ones are trained with 512p to save vram Model can now run longer and constantly till 30+ seconds. Way longer than without Lora The loras strength is still clearly T2V but it can also do I2V (reference model is not tested yet but be curious to do that and show results) for workflows i used the standard one nothing special you can find them in the galley files i attach to the lora. Even its V2 now the lora is Highly Experimental so don't expect that all runs flawless yet What has changed: - Trained on way bigger database - including better sound and lip sync - more styles ( comic , anime, realistic) and way scenarios - more specim - trained in music from my ai music side account so enjoy better music too :-) Settings: i use a weigh between 0.5 and 0.9max! it works with the turbo loras i use mostly 10 steps i tested res multistep/simple and euler/beta both works Important: if you wanna support my work please give alike it helps alot. also i like when people post what they do with it. have fun with the lora! ─────────────────────────────────────────────────────────────────────────────── EXAMPLE MEDIA ─────────────────────────────────────────────────────────────────────────────── 6 preview(s) in ./example_images/ — community results for this LoRA. Videos keep a matching _poster.jpg still frame. No prompts are recorded below: CivitAI's public API flags these posts as having generation metadata but does not return it, so the prompts cannot be scraped. Open the model page to read them. 01_video.mp4 (video, 1152x640) still: 01_poster.jpg 02_video.mp4 (video, 1280x704) still: 02_poster.jpg 03_video.mp4 (video, 1056x704) still: 03_poster.jpg 04_video.mp4 (video, 1280x704) still: 04_poster.jpg 05_video.mp4 (video, 736x1088) still: 05_poster.jpg 06_video.mp4 (video, 1056x704) still: 06_poster.jpg ─────────────────────────────────────────────────────────────────────────────── HOW TO USE (COMFYUI) ─────────────────────────────────────────────────────────────────────────────── 1. Copy the .safetensors file into ComfyUI/models/loras/ 2. Load the matching H3 checkpoint with UNETLoader. 3. Insert LoraLoaderModelOnly between the loader and your sampler. 4. Set strength as listed above, then add any trigger word to the prompt. 5. Keep CFG at 1.0 with BasicGuider unless you know you want real CFG. ─────────────────────────────────────────────────────────────────────────────── Scraped by lorakit on 2026-09-04 23:06 UTC. Stats and description are the author's; pros/cons are derived from the evidence tagged beside each line. ───────────────────────────────────────────────────────────────────────────────