Transform any item into a humanoid robot.
example:
BXJG method,
This video shows the process of transforming a pink Porsche 911 sports car into a humanoid robot in the style of a Porsche 911 using the BXJG method.
……
Finally, the pink Porsche 911-style humanoid robot appeared. It stood tall with an heroic posture, looking up at the sky.
You can add some descriptions, such as "This robot glides on the highway, its metal mechanical feet rubbing fiercely against the ground, generating some sparks"; "After the robot stands still, its right arm turns into a cannon with a golden lens as the muzzle, which it points towards the camera."
I'm not very good at this kind of description, and the effect of my own description is not very good. I would appreciate it if you, "experts in prompt words", could take care of it and upload your excellent works for us to learn from.
Thank you.
Description
This is a high-noise LoRa, which is only loaded in the high-noise sections. The low-noise sections do not require loading this LoRa.
FAQ
Comments (5)
Hello, that is nice idea. Thanks for that.
But I have a Issue. I use Low and High Noice. I Use your and other descriptions. But the failed because the “transformation” looks like a dark, smoky/pixelated cloud instead of crisp mechanical parts. There are no recognizable car features (front, wheels, body lines) and no sharp panel edges, joints, or distinct rotating/separating components. As a result, the robot silhouette isn’t clearly defined—it's mostly a blurry shadow with lost detail.
You are have idea what I can make it wrong?.
I use WAN2.2 14B I2V
Hello. This is a LoRa trained using a high-noise model. Please do not load it into the low-noise part. Here is an example of dataset annotation for your reference. I hope it can help you:
BXJG method,
The video shows the process of transforming a yellow Chevrolet Camaro car into a yellow Chevrolet Camaro-style humanoid robot using the BXJG method:
The car chassis folds upward, the hood lifts to form the chest plate, the doors swing out to become arms, the wheels rotate and detach to form legs, and the joints lock with hydraulic clamps;
Component mapping: Front hood -> Chest plate (upward flip), Side doors -> Upper arms (outward swing), Front wheels -> Knee armor (reassembly), Side mirrors -> Shoulder pads (folding inward), Grille -> Jawline (compression), Headlights -> Eye sockets (repositioning);
Style inheritance: Primary yellow from car body -> torso and limbs, black trim -> joint seams and armor edges, metallic gray -> internal chassis and faceplate, green accents -> head optics and sensor housings;
Mechanical details: Hydraulic pistons extend and retract, gear teeth mesh precisely at joint interfaces, sliding armor plates glide along track rails, energy glow pulses through circuitry pathways, servo motors engage for final locking, and visible torque rods stabilize limb articulation;
Finally, a yellow Chevrolet Camaro-style humanoid robot emerged.
Took me a while to get it to work I also was getting the smokey/pixelated cloud, Since I didn't know which place to put the lora, I have it in both with high being .75 or so and lower at a real low number like .3. This would align with the author response of it only needs to be in High Lora.
Also the Wan2.2 I originally was using was Smoothmix 2.2 which is for NSFW stuff and didn't work very well, so tried the official Wan2.2 FP16 which works but had some small issues, then I tried Wan Remix which is what I showed for the Ford Fiesta showpiece in the gallery.
Spent like a good 4hrs trying to get it right lol. Also you have to be detailed in how you want it to transform or you'll just get anything the AI wants to do. Like people getting out of the car and then it transforms or other objects entering the scene.
@EbonEagle Yes, when loading Lora into the high-noise part, setting the weight to 1 is sufficient because it is an action model. Therefore, I only trained the high-noise Lora, which does not participate in the sampling denoising of the low-noise part. The training dataset for the Wan 2.2 14B MoE model has a maximum of only 24fps, which is barely usable for this kind of "transformers" action. My dataset consists of 120 frames at 5 seconds, which were annotated using Qwen VL3 after frame extraction. I only trained for 2400 steps, and the results are not very good. Regarding how to describe the images specifically, you can try more and if you have good works, please share them with me.
@assyaya Hi many, thanks. That was my issue. I have High and Low active. I not know that shoudl not use on the LOW. Thank you.