All the stages suffer from similar problems.  Verticalfollow defined incorrectly, no floor tension (not that big a deal since the stages that do move vertically move very little), single layer-single image backgrounds (most of which should have been split into multiple layers), and varying degrees of color loss.  Player shadows should also be larger.

Dark Forest:
Stage is asymmetrical.  Boundleft should be the negative value of boundright.  Vertical follow should be higher.  The character is well off screen before the top of the stage is reached.  Try something like .3 or .4.  Try splitting the image into multiple layers and using differing delta values.

TMNT Alley:
Stage is asymmetrical.  Try multiple layers with differing delta values.  The floor looks awful.  The diagonal lines on the concrete change directions.  That should have been edited to look more consistent.

Gotham II:
Stage is asymmetrical.  The snow particles are too big for the most part.

Dragon Castle:
Missing verticalfollow.  Try .3 or .4.  The floor looks awful.  You can see that symmetry was used to extend the floor horizontally and vertically...not good.  Try a better edit or find a new floor.  The rain doesn't look good either (even after adding transparency tags).  The animation does not loop well.  I would get rid of it altogether, find a different rain graphic, and use velocity tags instead of animation to simulate movement.  The lighting effect on the floor is nice, however.

Gotham:
Stage is very asymmetrical.  Again, snow particles are too big.  The deltas are messed up, which makes the floor broken.  The floor layer always has a delta of 1,1 for low-res stages; no exceptions except for parallax (even then, the zoffset line should have an effective delta of 1,1).  Objects behind the floor should have a slower/smaller delta than 1,1.  Try 1,1 for the floor and .75,1 for the backdrop (left y-delta as 1 since the stage does not move vertically).  The backdrop does not tile well; needs further editing.  The backdrop should have also been split into multiple layers.

Kiss:
Stage is asymmetrical.  Needs verticalfollow.  The zoffsetline is too high (value too low); characters feet are too close to the top of the floor.  Try something like 225.

The color loss here is just horrendous.  Methods for correction:

a) Jigsaw method: most common method used.  Basically, a large image is split down into smaller parts, each part is separately color reduced to 256 colors, and then all the pieces are reassembled in the def file (like a jigsaw puzzle).  Each part doesn't have to be symmetrical (large pieces for less colorful parts, smaller pieces for more colorful parts), but most people prefer symmetrical parts (I've used 64x64 size tiles).

To absolutely avoid color loss, you would need 16x16 tiles (256 pixels max), but almost no one does that.  There are programs available which automatically split up images into small parts, but of course, none provide mugen start coordinates.

b) Dither method.  Sometimes, regardless of what's done, mugen will still auto reduce colors on its own giving less than satisfactory results (display output is locked to 16-bit color).  This is most notable when trying gradients.  For those, I take a screenshot, see what colors mugen allows for the gradient, and then make a dither using only those colors.  I try to avoid dither whenever possible, otherwise.

c) Modified transparency method.  You start off with 1 image and end up with 3.  Basically, an image is split into its individual color channels: red, green, blue.  Many programs can do this, but can't seem to set the other two channels to 0, however.  In order for this to work, the other two channels have to be set to 0 and not any other value.

After the images are split into channels, you should have 3 images: each one having shades of one primary color.  If not, then the other channels are not set to 0.  Color reduce each part to 256 colors, if not done already (there should be no color loss when doing this, since each channel only allows 256 values).  Use one channel layer as a base without transparency.  Then set the other two layers on top using the same start coordinates and with trans = add.  This will perform a simple layer addition and reproduce the original image.  I don't recommend this for most cases since it causes a rather significant performance hit for large images.  As a side note, this cannot be used to fix gradients; mugen will still auto reduce the final image on its own.

Kiss II:
Ok, this one has all sorts of issues.

Asymmetric floor.  The video screen area has ridiculous color loss.  Verticalfollow is way too high.  Try .6.

The deltas are a complete mess.  Everything is locked to 1,1.5...which breaks the floor as well.  Again, as a rule of thumb, objects behind the floor have deltas smaller than 1,1.  The floor and objects directly connected to the floor have a delta of 1,1.  Objects in front of the floor have deltas greater than 1,1.  If everything here was 1,1, it wouldn't have been so bad.

So, overall, everything needs quite a bit of work.